The AI Agent Swarm has come to Drupal Canvas! 🔗 **Source:** [The AI Agent Swarm has come to Drupal Canvas!](https://www.youtube.com/watch?v=AkQlvdm9VAY) > [!note]+ Document Notes > tags: [drupal, ai-agents, canvas-ai, agentic-search, drupal-ai] > author: Freely Give > source: DrupalCon Vienna > reference: https://www.youtube.com/watch?v=AkQlvdm9VAY > date created: 2024-10-16 12:00 > > # 🏷️ Title: The AI Agent Swarm has come to Drupal Canvas! > <!-- deep-read:sections start --> ## 🧰 Action Inbox - [H1] Install AI module and explore 22 available providers — DrupalCon Vienna demo #drupal-ai - [H2] Try Drupal CMS agents for agentic search capabilities — Canvas AI features #ai-search - [H3] Test Flow Drop module for visual AI orchestration — workflow builder demo #orchestration ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [Drupal AI Academy](https://drupal.org/ai-academy) - [Drupal AI Initiative](https://drupal.org/community/initiatives/ai) ## 📝 Summary - Jamie from Freely Give presents major advances in Drupal AI capabilities including Canvas AI, agentic search, and orchestration tools - Drupal AI Initiative has grown from 5 to 20 companies with 70 interested organizations contributing - Key innovations include automated landing page generation, AI-powered search with facets, and Flow Drop visual orchestration - Vision centers on human-AI collaboration with trustworthy governance and open source flexibility - Bottom line: Drupal's unique site builder concept and community collaboration positions it to compete with Silicon Valley AI startups ## 📖 Main Content The Drupal AI Initiative has evolved from a single-company project to a massive community effort spanning nearly 20 companies. The core AI module provides an abstraction layer supporting 22 AI providers, enabling flexible integration across different services. Canvas AI represents a breakthrough in automated content creation, capable of generating complete landing pages with appropriate images and content based on simple prompts. The system uses agentic search to find relevant media and creates contextually appropriate layouts. A major innovation is the Context Control Center, which provides centralized management of organizational context that AI agents can access when making decisions. This addresses the challenge of providing sufficient background information for AI to make informed choices about design, content, and functionality. The initiative has developed sophisticated search capabilities that surpass traditional filtered search interfaces. AI can interpret complex natural language queries like "upcoming non-European clinical trials for lung cancer" and automatically configure appropriate facets and filters, even when Google's search fails to understand the query. Flow Drop emerges as a visual orchestration tool that allows users to create AI workflows through drag-and-drop interfaces, combining deterministic steps with AI decision-making for more reliable automation. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=0) — [H1] Introduction and Energy Check - Key idea: Speaker acknowledges lower energy but commits to clear presentation - Evidence: "I've been told a lot I have a lot of energy for AI stuff and it's been a tiring couple of days" #### ⏱️ [00:02:15](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=135) — [H1] Drupal AI Initiative Growth - Key idea: Initiative expanded from 5 founding companies to nearly 20 with 70 interested organizations - Evidence: "We went from five founding companies to now almost 20 and apparently we've got 70 interested organizations" #### ⏱️ [00:03:30](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=210) — [H2] AI Module Statistics - Key idea: Version 1.2 shows massive community adoption with 1000 commits from 130 authors - Evidence: "The 1.2 branch of the AI module had a thousand commits with 130 authors and the downloads are accelerating" #### ⏱️ [00:05:00](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=300) — [H2] AI Search Breakthrough - Italy Query Example - Key idea: AI search finds content traditional search misses through semantic understanding - Evidence: "You search for Italy and this had gluten-free pizzas as a recipe but nothing appears. You go to AI and it says recipes inspired by Italy" #### ⏱️ [00:07:45](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=465) — [H1] Complex Query Success - Non-European Clinical Trials - Key idea: AI search succeeds where even Google fails on complex semantic queries - Evidence: "We tried the same thing in Google... Google the first result was something from Spain. So even with Google search and the AI capabilities they have this concept of non-European it couldn't understand it" #### ⏱️ [00:10:20](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=620) — [H2] Facet Automation Behind the Scenes - Key idea: AI automatically configures search facets based on natural language without user interaction - Evidence: "AI takes the natural language query and finds all the filters and facets and essentially fills them out for you" #### ⏱️ [00:12:30](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=750) — [H1] Agentic Search for Medical Research - Key idea: Multiple AI agents perform research tasks and synthesize findings without hallucinations - Evidence: "Agents are going away and doing a series of searches and then looking at the results and coming up with a series of answers" #### ⏱️ [00:15:10](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=910) — [H2] Canvas AI Landing Page Generation - Key idea: Complete landing pages generated automatically with appropriate images and layout - Evidence: "It's thinking. That word thinking is going to become important soon. It's building this template. It's doing a lot of planning" #### ⏱️ [00:18:00](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=1080) — [H1] Context Control Center Innovation - Key idea: Centralized organizational context management enables better AI decision-making - Evidence: "What we found is we created this context control center that could put in context. It would be centrally managed" #### ⏱️ [00:21:45](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=1305) — [H2] Component Metadata Schema - Key idea: Rich metadata for design components enables AI to understand when and how to use them - Evidence: "Best practices when to use it and when not to use it because you might start wanting to describe the goals and purpose of your page more" #### ⏱️ [00:25:30](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=1530) — [H3] Levels of Metadata Quality - Key idea: Bronze, silver, gold, and platinum tiers of metadata accuracy for different use cases - Evidence: "Having inaccurate metadata is actually a lot better than none when it comes to AI" #### ⏱️ [00:28:15](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=1695) — [H1] Philosophical Shift - From Thinking to Being - Key idea: Focus should shift from AI thinking processes to creating conscious entities with goals and memory - Evidence: "People have missed out on something by not philosophically focusing on trying to build a conscious being as opposed to just trying to do thinking" #### ⏱️ [00:32:00](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=1920) — [H1] Tools API Security by Design - Key idea: AI tools must integrate with Drupal's permission system to prevent security vulnerabilities - Evidence: "We need to make it secure by design. Every single tool will have permissions associated with it" #### ⏱️ [00:35:45](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=2145) — [H2] MCP Integration Strategy - Key idea: Model Context Protocol enables seamless integration with external AI ecosystems - Evidence: "If we make our tool as an MCP server work for MCP, then anyone who's made their tool have MCP clients can just use it" #### ⏱️ [00:39:30](https://www.youtube.com/watch?v=AkQlvdm9VAY?t=2370) — [H1] Flow Drop Visual Orchestration Demo - Key idea: Drag-and-drop workflow builder combines deterministic steps with AI decision-making - Evidence: "You can see here the nice flow of information and those tools... You can see a glance. And I think this is going to become key to doing AI orchestration well" ## 📚 References - [Drupal AI Module](https://drupal.org/project/ai) - [Flow Drop Module](https://drupal.org/project/flow) - [European Commission AI Hackathon](https://drupal.org/ai-academy) - [Drupal Starge](https://drupalstarge.com) ## 🗣️ Paraphrase - Drupal's AI capabilities have evolved from simple content generation to sophisticated agent orchestration, with nearly 20 companies contributing to a comprehensive ecosystem that includes automated page building, intelligent search, and visual workflow management - The breakthrough innovations center on context management and semantic search that outperforms traditional approaches, enabling AI to understand complex queries and automatically configure appropriate system responses without user intervention - The philosophical approach emphasizes creating conscious AI entities with memory, goals, and personality rather than just processing capabilities, positioning Drupal's human-centric design philosophy as a competitive advantage against Silicon Valley AI startups ## 📰 Newsletter The Drupal AI Initiative has reached a major milestone with Canvas AI now capable of generating complete landing pages automatically. What's most exciting is how this builds on Drupal's core strengths - the site builder concept that lets humans verify and refine AI-generated content. The agentic search capabilities are particularly impressive, handling complex queries like "non-European clinical trials" that even Google struggles with. The AI automatically configures facets and filters, making powerful search accessible through natural language. For developers, the new Flow Drop module offers visual orchestration of AI workflows, while the Tools API ensures security by design through Drupal's permission system. The Context Control Center provides centralized management of organizational knowledge that AI agents can access. Resources: [Drupal AI Academy](https://drupal.org/ai-academy), [AI Initiative](https://drupal.org/community/initiatives/ai), [Flow Module](https://drupal.org/project/flow) ## ✅ Action Items #### DevRel - Developer advocacy: Present Canvas AI capabilities at regional events, create video tutorials for agentic search implementation - Education: Develop curriculum for AI Academy hackathon, create office hours for Flow Drop adoption - Developer experience: Improve Tools API documentation, create MCP integration guides - Developer marketing: Launch campaign highlighting Drupal's AI orchestration advantages over competitors - Community management: Moderate AI initiative discussions, collect feedback on Canvas AI beta #### OSPO - Policy & Governance: Establish AI ethics guidelines for Drupal AI Initiative contributions - License Compliance: Review AI model integration licensing, ensure provider compatibility - Security Management: Implement security-by-design standards for AI tools, create vulnerability reporting process - Software Inventory: Track AI provider dependencies, maintain SBOM for AI module ecosystem - Developer Support: Provide training on secure AI tool development, create sandbox environments - Community Engagement: Coordinate funding for 70 interested organizations, establish contributor agreements - Strategic Alignment: Define KPIs for AI initiative success, align with Drupal Association goals #### Community - Maintainers: Support Canvas AI roadmap planning, establish governance for multi-company initiative - Contributors: Create first-issue programs for AI module development, onboard new contributors from 20+ companies - Sponsors: Develop transparent budgets for AI initiative, create impact reports for organizational sponsors - Ecosystem: Coordinate cross-project integration between AI module and Canvas, establish compatibility standards #### WordPress - Core AI: Study Drupal's agentic search implementation, evaluate Context Control Center concepts for WordPress - Community: Analyze Drupal's multi-company initiative model, consider similar collaborative approaches - Plugins: Research security-by-design approaches from Drupal's Tools API for WordPress AI plugins #### The WPCC - Programs: Develop AI-focused fellowship programs inspired by Drupal's collaborative model - Projects: Create governance frameworks for multi-stakeholder AI initiatives - Funding: Establish matched grants for AI accessibility and security research #### FAIR - Technical independence: Document Drupal's approach to AI provider abstraction, evaluate vendor lock-in prevention - Policies: Study transparent funding models from Drupal AI Initiative for community sustainability - Governance artifacts: Create templates based on Drupal's multi-company governance charter ## 🧪 Hypotheses - Visual orchestration tools like Flow Drop will become essential for managing complex AI workflows as deterministic steps outperform purely AI-driven orchestration - Context Control Centers will emerge as a critical architectural pattern for enterprise AI implementations requiring organizational knowledge management - Agentic search will replace traditional faceted search interfaces as natural language processing capabilities mature and user expectations evolve - Security-by-design AI tools integrated with existing permission systems will become the standard approach for preventing AI-related vulnerabilities ### 📊 Methodology - A/B testing of traditional vs agentic search interfaces measuring user success rates and satisfaction - Longitudinal analysis of Canvas AI generated content quality compared to human-created landing pages - Performance benchmarking of Flow Drop orchestrated workflows against code-based AI agent implementations ### ❓ Questions - How will the 20+ company governance model scale as the AI initiative grows to 70+ organizations? - Can Drupal's human-centric approach to AI maintain competitive advantage as other platforms adopt similar philosophies? - What are the long-term sustainability models for community-driven AI development competing against venture-funded startups? ### 🔮 Future Research - Investigation of consciousness-based AI architecture patterns for creating more goal-oriented agents - Development of standardized metadata schemas for cross-platform AI component interoperability - Exploration of federated AI model training using community-contributed datasets from multiple Drupal installations ### 💡 Implications - Practice: Site builders will need training on AI orchestration concepts and context management strategies - Policy: Organizations must establish governance frameworks for AI-generated content and automated decision-making - Security: AI tool permissions require integration with existing access control systems and regular security audits - Education: Curriculum must shift from teaching AI as a tool to teaching AI as collaborative agent architecture - Funding: Community-driven AI initiatives require new financial models to compete with venture-backed solutions ### 🧠 Learning Styles Assessment Reading level estimate: 12-14 (Graduate level due to technical AI concepts) #### 👀 Visual learners - Present: Flow Drop visual workflow diagrams, Canvas AI generation demonstrations - Improve: Architecture diagrams showing agent interaction patterns, visual metadata schema representations #### 🔊 Aural learners - Present: Conference presentation format with speaker explanations - Improve: Podcast series breaking down complex AI concepts, audio descriptions of visual workflows #### 🗣️ Verbal learners - Present: Clear definitions of agentic search, orchestration, and context management - Improve: Glossary of AI terminology, analogies comparing AI agents to human collaboration patterns #### 🧩 Kinesthetic learners - Present: Hands-on Flow Drop demo, Canvas AI interactive generation - Improve: Step-by-step tutorials for building AI agents, sandbox environments for experimentation #### 🧠 Logical learners - Present: Technical architecture explanations, security-by- <!-- deep-read:sections end --> ## 📺 Video ![](https://www.youtube.com/watch?v=AkQlvdm9VAY) --- # 📄 Full Transcript *Auto-generated from YouTube captions · timestamps roughly every 30 seconds · [watch on YouTube](https://www.youtube.com/watch?v=AkQlvdm9VAY)* **[00:00]** Hello everyone. Yeah, I uh I've been told a lot I have a lot of energy for AI stuff and uh it's been a tiring couple of days, so my energy is a little bit low, but I'm going to try my best to while still I usually have someone if somebody in the audience can just wave or yell if I start moving too quickly, talking too quickly, justiculating too widely, hard to understand, it would be really good. But um yeah has um so um uh is I mean **[00:33]** I'm guessing everyone here saw the Dre note. Um did anyone get surprised by anything that you saw in the Dre note to do with AI in canvas AI? Can we get any show of hands? Did anyone find anything there surprising? Uh so it was mostly what you were expecting. Cool. Um, my name is Jamie and I am the co-founder of a tech agency called Freely Give and uh, yeah, I've **[01:04]** been focusing on native Drupal CRM for a while and uh, I've recently become obsessed with AI and um, I'm one of the leads in the AI initiative, but I'm not a programmer. Um, could I get a show of hands? Who has heard me talk about AI before here? Okay, a lot of people, right? So, um I will not go to too much of the basics of like me and Drupal AI, but I want to say like I still feel like Drupal is uh surging forwards. The commu **[01:36]** the competition is getting a lot more fierce than it was like a year ago. People are stepping up, putting effort into it. Others are doing a good job at certain things. I still believe that Drupal is surging forwards. I still think it's never been a better time to do it with Drupal. Um, who here is interested in philosophy? Okay, who gets into like the weird kind of continental existentialism style philosophy? I mean, Vienna is maybe a sort of good place for this. **[02:08]** Yeah, a little bit. Okay. Well, we'll get a little bit into philosophy for and and we're gonna with Adrio, it's he's a really great storyteller. So, he had a story about everything. Um, what I'm going to do is go into a little bit of detail about what we've done and just show you things and a little bit about where we are going or could go. Um, so yeah. Yeah. This was where where I was a year ago, almost a year ago. Uh, there was a hackathon at the European Commission. **[02:39]** Um, a lot of cool energy started there. October the 16th, which is uh my birthday. So, it's also my birth. Yep. It's also my birthday tomorrow. So, again, I'm away during my birthday on Drupal AI stuff, which is ob obviously where I love to be and what I love to do. Um, but uh yeah, there's going to be another hackathon January the 27th and 28th. Uh, so this is the AI module. I'll move past this. Obviously, the core of the AI module has been **[03:10]** abstraction layer with 22 providers. Um, that's been pretty good. we've got some ideas of how we can take this further and make it scale more. Um and then um what we moved to was from the AI module to Drupal CMS agents which um has everyone tried out the Drupal CMS agents here? Yeah. Cool. And I think yeah, we shocked people a bit because at the time we decided to go with AI that could actually do things rather than AI that merely explained stuff or generated **[03:42]** content. And um we're still going there. So what are we doing next? Well, the AI initiative was launched. And has everyone here heard of the AI initiative by now? Yeah. Great. So when we started this, the AI module was originally built by not just one company. A lot of initiatives in Drupal tend to be owned by a company. That company gets a return on investment by being the deacto standard at doing that thing. And from the beginning we wanted **[04:13]** we thought AI was going to be bigger than a single organization. And so the AI module itself started with a collaboration with multiple companies and that's grown to an AI initiative with an official body to manage the fact that multiple organizations are giving multiple FTEEs. Um we have AI human collaboration as at our heart trustworthy governance open source flexibility and community innovation. and talk a bit more about that. But um um we had to grow to capture all of this **[04:44]** energy, all of these potential. Uh so we went from five founding companies to now almost 20 and apparently we've got 70 uh interested organizations. Uh the 1.2 branch of the AI module had a thousand commits with 130 authors and the downloads are accelerating. We're on about 500 a week, so up to 8,000 downloads. This was a little view of a potential road map of everything that we could start doing and start building. **[05:14]** So in terms of what we've done, I want to give a couple of little practical examples of things that have happened. Uh go into a bit more detail about the things you saw on the Dre note and I want to talk a little bit about what I think the future could be for us between now and Chicago. So AI search um [clears throat] one thing we did working with drop solid and the European Commission is we've created on it's on Drupal.org you can go to the AI Academy and it's a series of talks um helping you get into knowing how to use **[05:46]** Drupal AI with the goal being this uh hackathon in January and uh for the search one that we did recently. Uh yeah, you can see there's a playlist some of those you can subscribe to uh to get notifications for and you can also go through the drupal.org and some are on the on the web now. You can you can watch them. And um one thing we've started doing a lot more is stuff with AI search. We're finding it's it's really getting to a place that it's useful now. And so this was a simple example of the early days of using AI search to talk your data. **[06:19]** Uh and we had like an example in Omi where you search for Italy and this had gluten-free pizzas as a recipe but nothing appears. You go to AI and it says recipes uh inspired by Italy and you can see a deep Mediterranean kiche and a gluten-free pizza appear. And so we saw really early on AI search could find things that traditional search couldn't couldn't do very easily. But also there was some negatives. It was some things traditional search was good at. AI search wasn't there for. And then **[06:50]** there was the warrior of hallucinations. So with AI search that allows you to take your content and bring it into the context into the prompt that you send to AI and so AI when it gives you a response it's less likely to do hallucinations. Um it's more likely to find that content and find the information based around an understanding of what you've said. Um we made it work with search API. It's very pluggable. But here was a really nice simple feature. **[07:22]** So, who here is familiar with websites like this with a huge number of filters and facets on the left? Okay. Now, who enjoys using this? Okay, we got a few hands, right? [clears throat] What we found is that this is common in Drupal. Um uh especially if you have any client websites who have sites like this uh and you've been um an included or or anthropic or the AI crawlers came along along you you suddenly realized that you **[07:53]** had site sites like this because of the way views filters works with a new URL. Um AI loved going through every single combination of filters. But um what we had here is an example of something is aimed at technical users where we can see a series of studies uh sort of yeah studies about um uh various different kinds of studies going on across the EU and etc. Uh this is an it's all publicly available data from one of the uh European Commission websites and it's **[08:23]** for technical people so they can click these filters to find what they want a lot of the time. But we had this one which was upcoming non-European clinical trials for lung cancer. And you can see if I search for that the first thing is I get something is is Germany, Spain, United Kingdom, United States. The next one is all Europe. Uh so it didn't work. But um interestingly, we now do this and we've got AI behind the scenes and we've got the right answer. And um we actually found we **[08:54]** tried the same thing in Google. So we typed the same thing into Google by doing site search of that page and Google the first result was something from Spain. So even with Google search and the AI capabilities they have this concept of non-European it couldn't understand it. But why does this work? Well, behind the scenes, we have a very simple feature where AI takes the natural language query and finds all the f uh the filters and facets and essentially fills them out for you and redirects you to a page with all those **[09:25]** filters filled out. um you can choose to uh not expose that to end users like on a previous site and it just works like magic or here we've got AI writting out why it believes it picked those filters and even give the end user some ways they could refine their search and you can see immediately it has China at the top United States at the bottom and then you see Japan it's so powerful it's so simple there's no chance of hallucinations at all because the fi I mean if you don't give that block that **[09:55]** explains it AI is not writing anything out to the end user and um uh it just works. It's a really big powerful place where because Drupal is already there with a lot of people by doing relatively small but powerful advancements in AI we can just immediately make a difference to the problems that people have. So we went a bit further with a Gentic search. So, here's a whole series of uh COVID related papers um uh that that we found, a collection of them. And we're **[10:27]** we're we're pretending we're a doctor. A patient has come along with these symptoms and we want a management plan. And we've got a whole bunch of management plan things in our in our system. And medical advice with AI is very scary. Uh you know, really early on I had a friend whose wife was a doctor and had a patient come in and said like she was a pregnant woman. She goes, "Well, I've started smoking just as you recommended." Like, what? What? Why? Well, I went to Google and the first thing at the top was that pregnant women should smoke. And it was a hallucination by AI because AI at the top was **[10:59]** summarizing everything. Um, and so it was it's very scary to do that. But what we have here is instead uh this is the agent explorer. What's hide behind the scenes is agents are going away and doing a series of searches and then looking at the results and coming up with a series of answers of what it thinks is relevant questions. Now, it could expose the end user a view that just shows you the answers. So, no hallucinations or we could go in and here AI has produced a summary of why that thing is useful. Uh, has anyone **[11:30]** ever used Claude uh Claude like um the desktop app or the the app at all? Yeah. Has anyone tried doing anything with artifacts? Cool. I'll show you an artifact I made for this talk later on. But with artifacts, instead of AI just producing the output, it will have like a let's just say like a Google doc or like a plan. And every time you're asking do something, it writes this artifact. And in the same same way with search, we can do the same thing where it's creating a report for your answer, but all these agents are going away and doing bit of **[12:00]** research and then refining its final result. Um, and this is something again that we're finding quite powerful with AI search. So, we can use AI and reasoning to look at the results. Uh, agentic search can actually explore if it's relevant and present to the user the content with no hallucinations. Um, can also provide the user it is reasoning. One thing that I found very powerful about has anyone tried perplexity or chatgbt web search at all? Yeah. So a lot of people I found like in **[12:31]** the early days we had to do things with like vector databases and completely new ways of hosting things and indexing things and those are still relevant and important but with the agentic search um that existed uh they just use normal search and it could still produce pretty great results. I use chatdy web search a lot. I still use Google but I use it a lot and it finds things that Google can't find a lot of the time. Now, if we bring it into Google, if I actually tried that uh the non-cancer the cancer study one in chatb search and it still **[13:02]** couldn't get a right answer. So, it was still even more powerful to do it in Drupal because we could provide the AI agent specific metadata and guidance to help it more likely find those kinds of queries. Um, yeah, and there's other things that we've done. So in the dream note you saw that we built a whole page and we we found images and we put those images in the landing page but out of that we have some immediate impact stuff. So here's a **[13:33]** recipe that was produced by the immediate impact tract with uh the AI initiative and it's an image classification because in order to do search for images better what we found is we need to generate a lot more rich metadata. Normal search you want to do really finely tuned clear metadata like uh really clear tags and with humans you don't want too much noise on the page but with AI you want detail you want use cases for the module for example you want um descriptions of the image but **[14:03]** how to use the images how does that image make people feel what are the marketing goals you want to bring as much of this stuff into your system so that when search uh AI search goes along it's more likely to hit and so we've We've got a um we've got some recipes now that even if you don't use canvas yet, you can still use some of the benefits of AI search for finding media. So the landing page builder, you already saw it in the DE note, but here's another little video of everything **[14:34]** working for a slightly different case. Um so we have here it's thinking. That word thinking is going to become important soon. Uh, it's building this template. It's doing a lot of planning. And then we're going to see, uh, that page get streamed in. I just love watching AI do its little thing. Um, uh, and you can see the data, the the the information on the right hand side that it's being filled out. All that thing is there. And we're now getting to a place **[15:05]** where some pretty nicel looking pages are are being made. I mean, can we get a clap for this? Well done, Canvas AI. [applause] As you heard, one thing that came out of this, it wasn't part of our initial plan, but it became a n natural thing that we found become even more important is um if we're going to make it make good pages, you actually need to know a lot of context. And because a lot of the time when you build a web page you you know when you when you know let's say a **[15:35]** front-end developer gets started and starts building a landing page a lot of conversations have happened that it brings in a lot of context that you might take for granted into your into that person's brain information about the organization how they feel um does anyone know about hack and news the website the Y combinator one right so if you look at that design does anyone think it's pretty Right. I remember hearing one person who was like a you know relatively junior in design and this person believed that **[16:05]** every single page web page needs nice animations to look good. I was like well here's hacken news. This has nothing but all of the world's money kind of flows through this place. If you can get anything on the front page of hacker news then just loads of potential jobs uh uh you know people who might hire you in Silicon Valley will just know about you. Um but it's very simple and so depending on your target audience, the personas, the goals you have, you might want a very wide variety of different outcomes. Uh and so what we found is we created this um context uh control **[16:37]** center that could put in context. It would be centrally managed by sometimes maybe a different team who are building landing pages and AI could make use of it. So here it's a description of your company. Um, and this flowed naturally out of trying to do things that were harder. Getting AI to build a landing page was quite complicated, but naturally out of trying to solve these hard problems, other innovations came out. And we think this is going to be really important for everything. I think one thing that will happen, maybe Drupal CMS 2.0 or maybe 3.0 is we probably want **[17:09]** every page to have a field that is hidden that is the purpose of that page. So that's a context for that page. This context is the sitewide one. but actually settings so that AI can be forced to know about the context, can choose it when it needs to. Uh it can be overridden in different situations. Um so I think um this is really important. Another thing that came out of it was uh more rich metadata for the design system. Uh this is not yet in canvas AI but it should be soon. And what we found **[17:41]** really early on uh with our demos last year is we had no metadata about the components. So when AI would build the page, it couldn't really understand what these components were. Um it was usually like a title describe them, but they were very poor. If we did the tiniest amount of metadata, a title and simple description, immediately the results were significantly better. But we also want to go further and and here uh with Salsa Digital um they've produced a metadata schema with some example **[18:12]** content and what it does is in every single component we have the simple metadata but also the advanced metadata and AI depending on how it's trying to solve the problem it can see a list of all the components with the simple things once it shortlists its components it can then get pull in more interesting metadata and information about it so it knows how to use it. Uh this will be things like its accessibility. It will be things like more details of exactly what the what the component looks like and how it **[18:43]** works. But this one's interesting which is best practices when to use it and when not to use it because you might start wanting to describe the goals and purpose of your page more. And so rather than AI looking at well the user has described I want a hero section which means and then they describe a hero section is like it's got some buttons it's got a call to action it's got all those things instead if it describes the goal we want these components to actually reflect why why would you use these and not other things in fact we probably want to go this f go further when it comes to design is bring more **[19:14]** and more of this rich metadata into what we do so that when AI is finding things it can look through things from the end user's goals point of view rather than a clear description of what it's building. Um, we've kind of been playing with this idea of levels of metadata. So, you got like the bronze one where you have basic title and description. It will kind of work. You could use uh we can it hasn't been built yet. We got the outlines of it, but we could use Drush to make it so it can like look at a component and then **[19:45]** fill out all the metadata using that schema for you. Um, having inaccurate metadata is is actually a lot better than none when it comes to AI. Um, sometimes completely wrong information can actually make it have more the more likely get the right answer. It sort of like hits the right bits of its brain in some way. Um, and then gold would be a human has verified it. And then we're we're exploring like a platinum thing where you got an externally verified metadata by experts like looking at those things making sure all the metadata about use, use cases, **[20:16]** accessibility is is seeped in like real research and this helps us move away from AI slop where it's sort of generating whatever it wants. Uh, you're going to see a vibecoded page I made called Drupal Starge. I thought it looked really nice. uh I was very happy with it and then somebody else built a vibe coded tool for analyzing the accessibility of local councils and it looked identical to mine and it was like more and more I'm seeing these AI generated things have a very similar look um because it's just picking sort **[20:48]** of similar JavaScript libraries and similar designs and so you need to go a bit further if you want to make something that is still special and has a human touch even if AI is helping. So yeah, I want to thank there was a whole bunch of people working on that, but there's a couple of people who really got involved in this stuff. Yeah, Akil. Yeah, thanks for Ahmed and Kristen from Salsa with the context control center. Um, we had this simple idea for the for the context **[21:18]** control center where it would just be a piece of text. It just be like a text file and he went away on a weekend and produced what he did and it was so much better than we thought we would get by now. Uh so it was really impressive. Narendra from Aquia. Um I've been really excited how much the community has upskilled. Uh this is no longer just um you know market although it was wider the AI module was wider than just one company. Marcus was doing a pred predominant amount of the AI work for quite a while. And the reality is now we have so many amazing contributors in the **[21:50]** Drupal community who really get AI, really get like hard things like the context management, uh, prompt management, um, short-term memory artifact, like all of these things they're getting and contributing. So, it's really growing and becoming a platform that I think you can trust. So, where are we going to go next? Um, I think a big thing is design migration. We're really close there. But if we can have an AI powered design system um that has sort of variables exposed in a database, AI can write it. I think **[22:22]** it'll be you can get AI to completely vibe code the entire thing. I can write themes for you from scratch. But this kind of vibe config approach has some benefits to be something that you could give to somebody who is an end user who knows very little about code and are very unlikely to be able to do anything that will break long like it will break short term much. should look terrible but it's not going to introduce security faults. So I think there's quite a lot we could do there with sign migration. I think that's something we can do soon. Um uh with Figma to Drupal we saw in the dre note uh we're going to work on with **[22:53]** with um there to bring that into um the canvas AI Drupal CMS stuff uh because I think that is still going to be quite powerful and I think we're going to start focusing a lot on the component creation uh side of things which is kind of linked to design migration and Figma to Drupal. So time to get philosophical. I think therefore I am. Um yeah. So one of the things that I've been fighting with some of my friends about is like it was for a while it was like can AI really reason? What does it **[23:25]** mean to reason? Then is can AI think? Is AI conscious? And um what has struck me is when you look at I think therefore I am the the the developer world is so focused on processes and thinking is a process. It's a way of doing things. It's a way your brain works and it's what developers like to to do is think loads and loads about processes but not necessarily why that process is happening or the goals. It's to make those processes better. And the AI community has done something similar where when they talk about intelligence **[23:56]** they talk about thinking a lot. Um but a lot of the time when you have that phrase I think therefore I am it starts an assumption there is a thing that thinks and um I think there's an element where people have missed out on something by not philosophically focusing on trying to build a conscious being as opposed to just trying to do thinking and that seems silly and philosophical and maybe it's just gibberish but one thing that occurred to me is really quickly AI was **[24:26]** producing so many who hears use clawed code. Okay. And how much time do you spend thinking about your claw MD file? Like do you see it? It's like I don't know what it seems like is when you start using claw code it seems magical. If you want it to take it further really quickly you start finding context management. The clawed MD file becomes your primary thing that can make you a winner compared to everyone else. Can you make it sing compared to it kind of working most of the time? And there's some sense in which the Claude MD file **[24:58]** is analogous to the thinking being. It's the thing that does the thinking. It's the thing that has the goals. It's the thing that has the purpose in it. Um so like I think there's an element where we need to start thinking a lot about consciousness and and creating conscious beings. Uh things like memory, goals, the conscious entity itself, a personality, uh learning. There's lots of weird things actual animals do. One thing my friend mentioned was like if you see rabbits jumping around uh trying to get away from something, they seem to **[25:29]** move in a random direction. And actually their brains truly work in that way. When they go into fight or flight response is what I've been told and they're scared of something. There's this like thing in their brain that switches on that just makes them move randomly. So they're not thinking I need to move here, I need to move there. They're not planning a path. They're almost their conscious brain is switched off and there's this new process that's happening. and um the conscious and and and and so there's there's some sense in which in our human brains we have bits of our brains left and right that we we **[26:00]** our brains put a lot of effort to actually try and make them not cohhere too much. Um we have personalities that have different fields on different days and I think that this is actually a very practical thing that we can think about to make our AI stuff better today because the AI companies aren't thinking about this very much. you know, with all the money they have, with all the effort they've done, the Claude MD file, a flat file was all they came up with. We're seeing people in the community create better management of contexts, better claw MD file management, collections of **[26:31]** sub aents that can do various things with knowledge. But I think it's something that we in Drupal have the power to do. It's us to like create the um the conscious being that has the goals, purposes, the end points um in it that then AI and the thinking can do. So this leads on to orchestration. I've seen this before. Not everyone agrees with me about this, but everyone is trying to see themselves as digital ops orchestration and I want Drupal to be doing everything do it with Drupal. **[27:02]** Um um you know I've I won't go into detail of this. I've talked about what an agent is, instructions, tools and swarm but um what I want to do is is build up a picture for you of what is important to doing orchestration well and creating that conscious being and creating those things and we have to go a bit into those some of those concepts again. So we've got tools here is a p important part of what makes an AI agent do things and we had a very simple one in 1.0 O of uh Drupal CMS. What we've got now is a **[27:34]** lot better. I'm seeing a lot more people who aren't Marcus be able to build agents successfully. Um but Michael Lander um let me go oh okay I think a slide about him comes up soon. uh there's a tools module now where he has been thinking about like what if we made the actions module in Drupal so actions plus like correct like what if we had it so that the inputs were structured the inputs had a title description and examples like one thing I've heard is a lot of the time when you have an action in Drupal and you're **[28:05]** trying to code for it and it's like the inputs like an entity is that sometimes it needs a machine name sometimes it needs an ID and you kind of have to like guess and try and see what works and if everything crashes you know you got the wrong So there's a lot we could do to make the inputs and outputs of the actions that we have in Drupal better for developers. If we if you can if you can look at action and know when I see this action I know the kind of structure of the output it will give I know what kinds of things it does. Is it a string? Is it text? Is it is it entities? Is it ids? Is it you **[28:37]** know what is it? Uh that will be great for developers but also it will help AI understand what it can put in and out. It will be helpful for visual flow builders which I'm going to come into onto views bulk operation MCP endpoints. So a big thing that I think we're going to want to put a lot more effort into is is really is is building out this tools API uh having it secure by design. So we have a fantastic tool that was uh really powerful. We made it a few months ago that could edit and create any content. **[29:09]** So you could do content variations, you could do content classification, you could do anything. But user entities are sort of content entities. So you could also create an admin account, change the password for any entity. And yes, we built constraints so you could make it secure, but by default you had to know to do that. And so we never released it because we're so scared like if we sent it out there, would somebody use this cool little thing to create variants of a blog post? then somebody would prompt inject um something to the AI to make it **[29:40]** so it takes over the whole site. So we need to make it secure by design. When you create these things, it deeply is linked to Drupal's permission systems. Um uh we're going to have AI agents all have roles assigned to them and maybe even users, but certainly roles. So every single tool will have permissions associated with it. Every single agent will and then also the user of those agents, the chatbot. So we need to have that by by design and so I think it's going to be really important to help us do this **[30:10]** orchestration. Well, MCP has everyone heard about MCP here. Yeah. The general thing here is uh before MCP you needed an API that would connect to an LLM. Now we can just do everything in theory via this MCP. If we make something work for MCP then it will work. If you make your tool as an MCP server work for MCP, then anyone who's made their tool have MCP clients can just use it. We've seen real promise there. A big gate to like why some people don't want to do everything with **[30:40]** Drupal and they want to use some other tool will be they already have loads of integrations. Well, if we can make our tools better at pushing back to an MCP endpoint because they're more well- definfined and we can make an MCP client in Drupal more effective, it means that we can be part of an ecosystem without too much extra work because we've done our APIs and the documentation around them at the API endpoint well enough. So, it can just fit in with this ecosystem. It means if you believe that Drupal won't do everything, then Drupal will sit in an **[31:12]** ecosystem you already have nicely. It also means that because it can sit into it nicely and Drupal's I think better than a lot of things, we can then start slurping things into Drupal more. Um I won't show the video, but we had an example at Dev Days where um we integrated with Slack and it only took a couple of hours to build and no code where you did at Drupal bot ask it a question. There was an ECA model that could take that response, go and do a search on your Drupal site, summarize it as an LLM and reply reply in Slack. And **[31:44]** this was doing search connecting with Slack, no code and no custom endpoints. It had this MCP client thing. You just put in the URL of the MCP server. Immediately the tools are there with descriptions of how to do it that AI can use. And now it just works. So I'm going to come on to flowrop. Um, and now we're getting closer to, you know, where I really think I want to put a lot of the effort next. So, with ECA, you saw the UI there. Here's an example of how you could build an AI agent with **[32:17]** Flowdrop. And so, we've got something here that is using, it's using spelt um, uh, but it's more like a JavaScript framework for creating these nodes. You can see here the nice flow of information and those tools. There's different versions of this where there's a simpler view, there's a more complex one. You can see a glance. And I think this is going to become key uh to doing AI orchestration well. Um we're finding with a lot of our AI orchestration that we have to keep getting AI to do it. And **[32:48]** sometimes that's fantastic, but sometimes you ask it to do like four or five steps and you have to put in like capital letters. You have to like threaten it. You have to remind it over and over again to do this thing next. Otherwise, it just forgets. And so what I think we're going to want is a mixture even in the core of the AI module that when we build AI agents, we have some steps that are deterministic using this kind of this kind of flow. And this will then help us create these tools that we can throw out to MCP endpoints, we can use AI agents, we can we can do things with **[33:20]** um we saw this with active pieces. Um and so you saw that there. And so there's a question is should we go to like something outside like active pieces um or should we do everything in Drupal? Well, we've got different approaches we could take. I've called it small no orchestration. So Drupal doesn't handle orchestration. Everything is all an external system. Drupal provides a endpoints through its maybe JSON API and something else handles everything. There's a medium in between **[33:50]** which is what you saw with active pieces uh in the Dreote where an external thing does the brunt you know a lot of the orchestration and integrates everything else but the internal process in Drupal are handled with the EC ECA or or some other uh uh orchestration and this is a mixture of the two and there's obviously what I love is large Drupal does everything [clears throat] Drupal can communicate with an existing enterprise ecosystem and just replace everything Um, what I think is one of the benefits of **[34:21]** this, okay, that came in the wrong place, this tools thing. One of the benefits is actually to some degree, whichever direction we go in, that the starting point is actually quite similar. We're probably going to start around the middle because we're going to need to provide better internal tools anyway, better MCP endpoints anyway for external systems to be able to use Drupal. Well, like we had a taxonomy agent in Adres note, and it didn't provide very good error handling. So it's a bit hard to understand why. That's because that taxonomy agent was built for our orchestration agent in Drupal CMS. It wasn't it wasn't made for **[34:52]** that. And whereas the the the active pieces went straight to it. And if we in in order to make it work well with other systems, we have to put a lot of effort into internal orchestration anyway. And so what that will do is if we want to do better at medium, we have to do a lot of the same things to get better at large. And then if we get really good at that stuff, uh what you might find is people just want to switch off their other things and just do more and more just with Drupal directly communicating with those systems. That was a tool. Um **[35:24]** yeah. Um and similarly what we find um when people have used NA10 in Drupal is there's a lot of complications when in dup in the Dre note you saw a very simple entity and you could map it but when you've got like paragraphs or you've got landing pages or you've got like really complex entities with lots of relationships it starts getting very hard to use your these visual things to click a source goes to all of these different places um especially if you're working with a database and you're just looking at Drupal's database instead of the JSON API it gets over overwhelming **[35:55]** But when we do a mixture of the two, we can do stuff to make it easier for you to have like a blog creating agent that is the input and whatever you throw at it, it can sort out where that goes. Um, logging I think is going to become very important and having consistent logging. If you have, you saw AI was making decisions and it was a very simple decision, but we've done claims management and if you're going to have AI make a decision on your claim, you're going to need to store that log for years. But a lot of the time when people **[36:27]** think logging, they think of like open telemetry and those logging systems that already exist. And is it aimed at developers or CIS admins who are coming into it every so often? But a lot of this logging needs to be understood by people with extreme domain knowledge on claims management, not developers to be able to go and see the flow of decisions. See what AI did and why it went wrong. And I think we're going to have to do I think it's going to be a lot easier to do that in Drupal than these other tools because I think a lot of them are going to struggle when they're focused entirely on some one **[36:57]** slice uh of of of logging. Um we also find our data normalization uh one of the reasons why Marcus came to to to uh to to build automators um instead of using lang chain was he was constantly having to write code to write plugins to to fit between things. Uh same thing with flowrop uh he tried to write plugins for flow-wise for with Drupal and it was just taking ages. It's just easier to do it in Drupal. So I'm going to move away now. We're going to have a brief thing from Shibin to take a look **[37:29]** at something with flowrop. >> Hello. Is it on? Do I have to turn it on? >> Yeah, there you go. >> Oh, good. Ah, thanks Jamie. I mean, this is how it feels to be on a big stage, huh? So, yes. So, what I'm going to five minutes. Yes, I'm going to go to it very fast. So, we have a module called flow. You can install it on your dup uh Drupal website. And I have here uh something which is already installed. And what it is is **[38:00]** like a workflow editor reimagined. Um so once you install it, you have to go through some like you know installation installation process. But you what you could do afterwards is like you can create a workflow. Okay. Uh I'll just go and open a one that is already filled with something. Yeah, demo sounds good. So let me just uh go through a small uh demo of like you know how I would see **[38:31]** like a conversational interface being made. So in which like you know you would start with having an input to capture the user's message that we could do from um pulling in a text input node and then we would use a chat model to analyze this message. Um and then we will connect the text input to chat model like this. And of course uh we need to show the message to the user. So what we can do is we can use um **[39:03]** a chat output node to drag and drop and then we connect the response to this. So this is like a very basic but what I will do is like now next uh we'll go into a little more complex workflow. Yes. And now uh and the point to note here is like yes we could think from a traditional point of view with no AI like you know how things get connected and like you know the functions gets called and all but with what what we can also do is leverage the metadata **[39:34]** provided by these tools and convert them into the tools which the AI can use because doing so me means less connections of course also and this is taking on like the tools API building upon tools API where the LLMs can figure out how to call the tools, when to call the tool and uh you don't need to like you know be connecting these individual things around but here um you have full control on like you know which tool calls what **[40:04]** and what data is available for each and for this I will go quickly jump to like a runtime context. Um so uh I'm using a decoupled API because this is all available as a decoupled API. So uh this is like a workflow editor in a decoupled mode but also this is a workflow which uh you can see a live demo on factorial if you want to factorial stand. Uh this is something which is connected to our forms where you enter the data use the choice and then for each new submission **[40:35]** that start it starts a pipeline and you can see uh these um as you submit uh this workflow gets triggered behind the scenes and it has like a couple of workflows here. The one on top is like uh deciding whether uh we should create an entry in the CRM based on the prompts provided by the sales representative um uh sales guys from factorial and in below like you know uh it is responding to conditional uhly responding to different uh workflows. For example, **[41:06]** there is this if else node here which decides if you have chosen AI or not and which path to choose because um yeah it will follow the true path if this condition is true and not. And this is not the AI path. This is like very much deterministic but we can mix this with an agent model. So we can very much control like when a certain agent gets fired and with what data because we have like you know two uh so because as you can see this pip uh this line controls what data is available to that. So we **[41:36]** have a global context but also this agent has a very small portion of the global context because we choose as creator of this workflow like what exact data is available to that. So the possibilities are like really awesome on this one and um also the real-time monitoring that happens. So these are the submissions which failed for example you get to see that something failed um because like these work these pipeline runs are stored in the database and each each job um what I call it like each **[42:07]** node has a job associated with it with its own state management and you can drill into like you know why it failed and all because it captures all the input and output it it got and created and here we can again also rerun this workflow again tomorrow if we needed and like because in this case we forgot the Open AI keys and so it didn't work. So yeah, that's it. [applause] Okay. Hello. Yeah, one of the reasons why I'm **[42:39]** so ill and tired at the moment is because I spent a lot of a lot of my last couple of nights just waking up dreaming about this. Literally just thinking about like I want to be using Flow Drop. I have dreams about me clicking around making my AI agents and all these things. And I'd wake up and think about like oh wouldn't it be cool if I had this sent should have been loads of messages of it needs this. So um I'm very excited by this and I think there is a sense in which this is a UI that end users can use but actually the important thing is constraining and **[43:09]** controlling the AI agents and the flow they do. So one use case that we're finding a lot of people have is they want to just do find and replace. So, they've got a whole bunch of content and there's like maybe a fact or statistic that's wrong. They found out it's wrong. And there's actually a university that was sued recently because somebody went to that university. Uh, and and what they had to do is they had to pay a whole team of humans to read through like 10,000 pieces of content to find any reference to this and change it. Um, now you could use AI to do that, but if you use AI to orchestrate looking **[43:40]** through 10,000 pieces of content, it just gets confused. It doesn't keep track of what it's doing. Um, it means that you have to like do a lot of context management with having one AI write out the answer. Like in our AI agents at the moment, sometimes when we need to pass chat history to another agent, we're getting one agent to like literally write out with output tokens that are slow and expensive. With flowrop, what we can do is build our flows where when we need the chat history for one, we can just pull it in as a thing as one of the inputs. So I **[44:10]** think there's going to be a lot that is going to be powerful for all AI agents that we use now anyway as well as using it for more deterministic workflows that ECA has been useful for um um uh that are less you know the less fluid types. Uh and then I think AI migration I think with AI migration when we started it we tried onetoone migration but that's rare [snorts] and so something with flowrop where you can just pull in a data source see all the possible things have AI blow **[44:41]** up unstructured data into different bits and then have something that maps it I think this is going to be really important there so um I don't have very much time now so Drupal Starge is something I've been vibe coding this is a place where we've got so many modules now across the AI ecosystem. So um this is um well if you go to it um you you what you'll see is that we try to collect everything that's happening all into one place keep track of what people are doing. Uh there's **[45:12]** actually road maps there. There's uh cambban board so you can see things that need reviewing. Um and we have something here that is like a product road map. Uh, and I sort of uh I did a vibe. This is like a, you know, claude code thing where I've got like you can see there's a lot that's been going on here. I haven't finished writing everything out, but we don't actually have the concept of deliverables really in Drupal, right? We we have modules, but those aren't a deliverable. Those usually are tools **[45:43]** used to build deliverables. We have issues, but issues are usually very technical. We have meta issues, but they can be anything, just a collection of issues that go on. So a deliverable of a clear thing that you can like use as a start and finish date we don't have. And so we've created our own metadata on drupal.org in the issue itself like in a little template to define something and then from that we can create a deliverable that then links to all of these issues and then actually keep track of what we need to do to do that. But there are things we have delivered. It's still hard to figure out how to **[46:13]** package it. Recipes is one way that you've seen people initiative have gone for and that will work for some of you. Sometimes you need a description of how to do it yourself. It's complicated what they are, but you can see that um like the AI agents testing framework is something where we've got some really cool tools where you can without code write out like a conversational history. Um and and then test the Drupal CMS agents and then keep trying them out. Um this will eventually help us I think fine-tune open source models. We've got the image classification. **[46:45]** Um, if if a AI agents can't find an image, it can pull it from pixels, which is a royalty-free site, so you can get images. There's a lot of stuff that is happening that we've done, and we're just trying to get better at showing you the stuff we're working on. And what I'm sort of going to start trying to get some help with um uh uh uh over the next couple of days is writing up some issues. Um uh so if you want to come on on a contrib like planning out these **[47:17]** particular things can we do uh okay I'm moving fast through this now these are places where you can go to actually try this stuff out for yourself but yeah these are kind of it's not set in stone yet but better focus on MCP AI migration but with like 80% savings making things 80% quicker rather than 100% with AI actions plus or tools API the visual flow builder stuff across the board UX improvements and agentic search. I think these are some things that are next and so uh I could do some help if anyone **[47:48]** wants to get involved. Before it was about just writing like trying to actually build things but I think now it would be good to do a little bit more planning than we've done in the past and so some help in like collecting those issues together putting structuring them having it something that that people can get going with. So I want to end this is what I said a long time ago. I still think it's true. Why do I That was an AI module. Um why do I think we're going to win? I think no one does humanentric like we do. The site builder concept is actually unique to Drupal. It's not really there in anything else. Um and so **[48:20]** if you need something where AI vibe code something, but then a non-developer can check it, Drupal's perfect for that. I think no one does open source like we do. Open source is usually one or two ginormous Silicon Valley corporations behind it. Um, we have a true ecosystem of open- source stuff. All of our modules are open source. They're not like a whole bunch of paid plugins. Um, and I think no one does community collaboration like we can do. Like freely give on their own, beating Lang Chain is impossible, but they have like **[48:52]** at the time they had 55 employees and we're getting close to that many FTEEs. So, across our community, I think we can beat some of these startups. Um, so thank you [applause]