Anthropic's Managed Agents Are Different (Here's Why)

Anthropic's Managed Agents Are Different (Here's Why) ![rw-book-cover](https://img.youtube.com/vi/uxHolzq1NcA/maxresdefault.jpg) 🔗 **Source:** [Anthropic's Managed Agents Are Different (Here's Why)](https://www.youtube.com/watch?v=uxHolzq1NcA) > [!note]+ Document Notes > --- > tags: #ai-agents #anthropic #claude #automation > author: [MISSING] > source: YouTube > reference: https://www.youtube.com/watch?v=uxHolzq1NcA > date created: 2024-12-19 10:30 > --- > # 🏷️ Title: Anthropic's Managed Agents Are Different (Here's Why) > ## 🧰 Action Inbox > - [H1] Test Claude managed agents for customer support automation — compare with self-hosted solutions #ai-agents > - [H2] Evaluate pricing model vs Claude Pro subscription for personal use cases — assess cost-effectiveness #pricing > - [H3] Explore MCP integration capabilities for existing workflows — investigate tool compatibility #integration > ## 🧭 Table of Contents > - [[#📝 Summary]] > - [[#📖 Main Content]] > - [[#🗒️ Notes]] > - [[#📚 References]] > - [[#🗣️ Paraphrase]] > - [[#📰 Newsletter]] > - [[#✅ Action Items]] > - [[#🧪 Hypotheses]] > - [[#📊 Methodology]] > - [[#❓ Questions]] > - [[#🔮 Future Research]] > - [[#💡 Implications]] > - [[#🧠 Learning Styles Assessment]] > - [[#📊 Accessibility Assessment]] > - [[#🌍 DEIB Assessment]] > - [[#✅ Logical Validity]] > - [[#🛡️ DARVO Analysis]] > - [[#🎯 Conclusion]] > ### 🔗 References > - [Claude Managed Agents Documentation](https://docs.anthropic.com/en/docs/build-with-claude/computer-use) > - [Anthropic API Pricing](https://www.anthropic.com/pricing) > ## 📝 Summary > - Anthropic launched Claude Managed Agents, allowing no-code agent creation on their managed infrastructure > - Architecture features session harness, orchestration, and secure credential management with runtime isolation > - Agents can integrate with tools like Slack, GitHub, and MCP servers without local setup > - Pricing model uses API tokens plus 8 cents per session hour, separate from Claude Pro subscriptions > - Positioned as "Apple-like" managed solution versus "Linux-like" self-hosted alternatives like OpenClaude > ## 📖 Main Content > Claude Managed Agents represents Anthropic's approach to democratizing AI agent creation through a fully managed platform. Unlike self-hosted solutions that require developers to handle infrastructure, security, and deployment, managed agents abstract these complexities behind a natural language interface and robust cloud architecture. > The platform's three-tier architecture includes a stateless harness (routing and execution), session management (persistent logging), and orchestration (failure recovery and scaling). This design enables automatic failover and horizontal scaling while maintaining security through runtime credential injection. > The service targets users who want production-ready agents without the operational overhead of self-hosting, though it comes with API-based pricing that doesn't leverage existing Claude subscriptions. > ## 🗒️ Notes > #### ⏱️ [00:00](https://youtu.be/uxHolzq1NcA?t=0) — [H1] Claude Managed Agents Introduction > - Key idea: New no-code agent creation platform on Anthropic's managed infrastructure > - Evidence: "Custom agents without writing a single line of code, and these agents run on Anthropic's own managed infrastructure" > #### ⏱️ [01:30](https://youtu.be/uxHolzq1NcA?t=90) — [H2] Platform Philosophy Comparison > - Key idea: Managed agents positioned as "Apple-like" vs "Linux-like" self-hosted solutions > - Evidence: "Open Claude is a bit like Linux... Whereas, Claude managed agents is, dare I say it, like Apple" > #### ⏱️ [02:45](https://youtu.be/uxHolzq1NcA?t=165) — [H3] Medical Agent Demo Setup > - Key idea: Real-world example using GitHub repo integration and Slack interface > - Evidence: "Private GitHub repo that contains all my medical information scraped from the NHS app" > #### ⏱️ [04:15](https://youtu.be/uxHolzq1NcA?t=255) — [H2] Agent Creation Process > - Key idea: Natural language setup through Claude console or TypeScript SDK > - Evidence: "Type here in natural language to create my agent from scratch" > #### ⏱️ [06:30](https://youtu.be/uxHolzq1NcA?t=390) — [H3] Tool Selection and Permissions > - Key idea: Guided tool recommendation based on agent purpose > - Evidence: "Recommends me to use the pre-built tool set, so read, glob, grep, and to not use write, edit, bash" > #### ⏱️ [08:00](https://youtu.be/uxHolzq1NcA?t=480) — [H2] Infrastructure Management > - Key idea: Agents deployed to Anthropic's managed environment with console visibility > - Evidence: "Inside the Claude console, I can see my agent over here, as well as the environment I've just created" > #### ⏱️ [09:45](https://youtu.be/uxHolzq1NcA?t=585) — [H1] Pricing Model Details > - Key idea: API token pricing plus session hours, separate from Claude subscriptions > - Evidence: "All tokens used by the managed agents are priced with the pricing model that the Claude API uses... 8 cents per session hour" > #### ⏱️ [11:20](https://youtu.be/uxHolzq1NcA?t=680) — [H3] Agent Performance Demo > - Key idea: Medical query demonstration showing tool usage and knowledge synthesis > - Evidence: "I should be cautious with calamari because I'm allergic to shrimp, which is true" > #### ⏱️ [13:15](https://youtu.be/uxHolzq1NcA?t=795) — [H1] Architecture Deep Dive > - Key idea: Three-component architecture with harness, session, and orchestration > - Evidence: "The harness, also known as the hands of the system... session here is like the memory... orchestrator is what decides what mode" > #### ⏱️ [15:30](https://youtu.be/uxHolzq1NcA?t=930) — [H2] Fault Tolerance and Scaling > - Key idea: Stateless harness enables automatic recovery and horizontal scaling > - Evidence: "Orchestrator can create a new harness, and because the session logs are separate... can read the logs to find out what's been going on" > #### ⏱️ [16:45](https://youtu.be/uxHolzq1NcA?t=1005) — [H1] Security Model > - Key idea: Runtime credential injection prevents model access to sensitive data > - Evidence: "Credentials are called at runtime... the harness or the model doesn't know anything about it" > #### ⏱️ [18:20](https://youtu.be/uxHolzq1NcA?t=1100) — [H2] Platform Advantages and Limitations > - Key idea: Easy creation vs limited customization compared to open alternatives > - Evidence: "Very easy to create an agent. You don't have to learn about SDKs... if you want to create something that doesn't exist within those bounds, then you just have to go ahead and write your own code" > ## 📚 References > - [Anthropic Managed Agents Documentation](https://docs.anthropic.com) > - [Claude API Pricing](https://www.anthropic.com/pricing) > - [TypeScript Claude SDK](https://github.com/anthropics/anthropic-sdk-typescript) > ## 🗣️ Paraphrase > - Anthropic's managed agents platform eliminates the need for developers to handle infrastructure, security, and deployment by providing a natural language interface for agent creation > - The three-tier architecture separates execution (harness), memory (sessions), and management (orchestration) to enable automatic scaling and fault recovery while maintaining security through runtime credential isolation > - While the platform significantly reduces complexity compared to self-hosted solutions, it requires separate API billing and may be limited for highly customized use cases > ## 📰 Newsletter > 🤖 **Claude Managed Agents: No-Code AI Automation is Here** > Anthropic just launched Claude Managed Agents, making it incredibly easy to create production-ready AI agents without any coding. Think of it as the "Apple" approach to AI agents - everything managed for you vs the "Linux" DIY approach of self-hosted solutions. > **What's exciting:** Create agents using plain English that can integrate with Slack, GitHub, and other tools. The architecture is impressive with automatic scaling, fault tolerance, and security through runtime credential injection. > **The catch:** You'll pay API rates (separate from Claude Pro) plus 8¢/session hour. But for businesses wanting AI automation without the infrastructure headaches, this could be a game-changer. > Ready to automate your workflows? The managed approach might be exactly what teams need to ship AI features fast. > Resources: [Anthropic Console](https://console.anthropic.com/), [API Docs](https://docs.anthropic.com/), [Pricing](https://www.anthropic.com/pricing) > ## ✅ Action Items > #### DevRel > - Developer advocacy: Create managed agents tutorials, comparison guides with self-hosted solutions > - Education: Workshop curriculum for no-code agent creation, best practices documentation > - Developer experience: Investigate SDK integration patterns, CLI tooling gaps > - Developer marketing: Case studies showcasing managed vs self-hosted trade-offs > - Community management: Gather feedback on pricing model, feature requests > #### OSPO > - Policy & Governance: Review vendor lock-in implications, data residency policies > - License Compliance: Assess third-party integrations, MCP server licensing > - Security Management: Evaluate credential management practices, runtime isolation > - Software Inventory: Track managed agent dependencies, API usage > - Developer Support: Training on managed vs self-hosted decision criteria > #### Community > - Contributors: Share managed agent templates, integration patterns > - Ecosystem: Cross-platform compatibility with existing AI tools > #### WordPress > - Core AI: Evaluate managed agents for WordPress admin automation > - Community: WordCamp demos of content management agents > - Support: Automated help desk agents using WordPress knowledge base > - Documentation: Agent-driven documentation updates, translation workflows > ## 🧪 Hypotheses > - Organizations will prefer managed agents over self-hosted solutions when infrastructure complexity outweighs customization needs > - The separation of subscription billing from API usage will create adoption friction for individual developers > - Runtime credential injection architecture will become standard for enterprise AI agent platforms > - No-code agent creation will democratize AI automation beyond technical teams > ### 📊 Methodology > - Survey developers comparing managed vs self-hosted agent deployment experiences and costs > - A/B test pricing models: subscription credits vs API billing for agent usage > - Benchmark security models: runtime injection vs embedded credentials across platforms > - User study: non-technical users creating agents through natural language vs traditional interfaces > ### ❓ Questions > - How does the 8¢/session hour pricing compare to self-hosting costs at different usage scales? > - What happens to agent availability and data access if Anthropic experiences outages? > - Can the managed agent architecture support real-time applications with low latency requirements? > ### 🔮 Future Research > - Integration patterns between managed agents and existing enterprise workflows > - Cost optimization strategies for high-volume managed agent deployments > - Security audit frameworks for runtime credential injection systems > ### 💡 Implications > - Practice: Shifts agent development from infrastructure management to business logic design > - Policy: Organizations need vendor dependency assessments for critical automation workflows > - Security: Runtime credential injection may become enterprise standard for AI applications > - Education: Training programs should cover managed vs self-hosted decision frameworks > - Funding: Budget allocation shifts from infrastructure to API consumption and session costs > ### 🧠 Learning Styles Assessment > Reading level estimate: Grade 12 > #### 👀 Visual learners > - Present: Architecture diagrams, console screenshots, demo flow > - Improve: Security model visualizations, cost comparison charts > #### 🔊 Aural learners > - Present: Video demonstration with clear narration > - Improve: Audio summary of key architectural concepts > #### 🗣️ Verbal learners > - Present: Clear analogies (Apple vs Linux), defined technical terms > - Improve: Glossary of agent terminology, more real-world examples > #### 🧩 Kinesthetic learners > - Present: Step-by-step agent creation walkthrough > - Improve: Hands-on workshop guide, interactive pricing calculator > #### 🧠 Logical learners > - Present: Architecture breakdown, pricing model explanation > - Improve: Decision tree for managed vs self-hosted choice > #### 🧑‍🤝‍🧑 Social learners > - Present: Comparison with community alternatives > - Improve: Discussion forum links, user testimonials > #### 🧘‍♀️ Solitary learners > - Present: Complete end-to-end example > - Improve: Self-assessment checklist for agent requirements > #### 🌿 Naturalistic learners > - Present: Medical data example with real-world context > - Improve: Sustainability impact of managed vs self-hosted infrastructure > ### 📊 Accessibility Assessment > #### Screen reader semantics > - Present: Clear heading structure, descriptive timestamps > - Improve: Alt text for architecture diagrams, ARIA labels for video controls > #### Keyboard navigation > - Present: Logical tab order through content sections > - Improve: Skip links for timestamp navigation, focus indicators > #### Contrast and readability > - Present: High contrast text, readable font sizes > - Improve: Code syntax highlighting contrast check > #### Text resizing and reflow > - Present: Responsive layout at 200% zoom > - Improve: Ensure timestamp links remain clickable when enlarged > #### Media and captions > - Present: Video with clear audio > - Improve: Full transcript, chapter markers for accessibility > #### Forms and interactions > - Present: Clear link descriptions > - Improve: Accessible video player controls > ### 🌍 DEIB Assessment > #### Cultural breadth > - Present: Technical content accessible globally > - Improve: Non-US healthcare examples, global pricing context > #### Inclusive language > - Present: Technical terms explained clearly > - Improve: Avoid assumptions about technical background > #### Neurodiversity support > - Present: Clear section breaks, consistent formatting > - Improve: Content warnings for complex technical concepts > #### Representation > - Present: Focus on technical merit > - Improve: Diverse use cases beyond medical applications > #### Psychological safety > - Present: Non-judgmental technical discussion > - Improve: Acknowledge different technical skill levels > ### ✅ Logical Validity > #### Strong reasoning > - Deductive strength: Clear cause-effect relationship between architecture design and scalability benefits > - Inductive strength: Single medical agent example generalizes to broader platform capabilities > - Causal reasoning: Runtime credential injection directly prevents security vulnerabilities > - Analogical strength: Apple vs Linux comparison effectively illustrates managed vs self-hosted trade-offs > #### Fallacies > - Cherry-picking: Only positive demo outcome shown, no failure cases discussed > - False dichotomy: Presents only managed vs self-hosted options, ignoring hybrid approaches > #### Tone of voice > - Overall tone: Enthusiastic but technically grounded, with practical cost considerations and limitations acknowledged > ### 🛡️ DARVO Analysis > - Deny: No denial patterns present > - Attack: No attacks on competing solutions beyond feature comparisons > - Reverse victim/offender: No victim/offender dynamics > - Emotional cues and rhetoric: Positive framing without manipulative language > ## 🎯 Conclusion > Claude Managed Agents represents a significant shift toward democratizing AI agent creation through managed infrastructure and natural language interfaces. The platform's architecture prioritizes security, scalability, and ease of use over customization, making it ideal for organizations that want to ship AI automation quickly without operational overhead. While the separate API pricing model and platform constraints may limit adoption for some use cases, the no-code approach opens agent development to non-technical teams and could accelerate enterprise AI automation adoption. > ## 🧱 Refactor Appendix > ### 📚 Literature Split Pad > Purpose: Stage CEE units that become Literature notes. > Split level: H3 > ### [H1] Managed AI agents reduce infrastructure complexity significantly > > [!quote] Anthropic's Managed Agents Are Different (Here's Why) by [MISSING] via YouTube. Published 2024. Accessed 2024-12-19. > > "Custom agents without writing a single line of code, and these agents run on Anthropic's own managed infrastructure." > Demonstrates shift from developer-operated to platform-operated AI agents. > [view highlight](https://youtu.be/uxHolzq1NcA?t=0#H1) #xref #articleseed ID:rw_managed_agents_complexity > ### [H2] Runtime credential injection prevents AI security vulnerabilities > > [!quote] Anthropic's Managed Agents Are Different (Here's Why) > > "The harness or the model doesn't know anything about it, and it can't have access to it." > Security through isolation rather than access control policies. > [view highlight](https://youtu. > --- > ## Full Transcript > [00:03] So, the Claude team have released Claude [00:05] managed agents, which is what they call [00:08] the next evolution after the agent SDK. [00:10] This is something that lets you create [00:12] custom agents without writing a single [00:15] line of code, and these agents run on [00:17] Anthropic's own managed infrastructure, [00:20] which has a very cool architecture that [00:22] is perfect for shipping secure, [00:24] scalable, production-ready agents that [00:26] can do pretty much anything you want, so [00:28] answer customer queries based on your [00:30] knowledge base, or even do regular [00:32] research for you and send it to you at a [00:34] specific time using Slack. But, why [00:36] would anyone use this instead of [00:37] installing Claude's code on their own [00:39] VPS? Hit subscribe, and let's get into [00:42] it. [00:45] But, before we get into it, let's talk a [00:47] bit about Anthropic because they've been [00:49] busy these last few weeks. I mean, they [00:51] recently prevented people from using the [00:53] Claude subscription on third-party tools [00:56] like Open Claude, which people think is [00:58] because of managed agents, but they're [01:00] not exactly the same thing. I mean, I [01:02] would say Open Claude is a bit like [01:04] Linux. It's a tinkerer's agent. You pick [01:06] your own hardware, you pick your own [01:08] model, you deal with security and [01:11] everything in between. Whereas, Claude [01:13] managed agents is, dare I say it, like [01:16] Apple because you don't need to do any [01:18] of those things. Claude takes care of [01:20] the infrastructure, the security. You [01:22] just tell it what you want in simple [01:24] English, and it will go ahead and build [01:27] your perfect agent. In fact, let me show [01:29] you how easy it is by going through a [01:31] simple, yet very personal example. I [01:34] have a private GitHub repo that contains [01:36] all my medical information scraped from [01:38] the NHS app, and I want to communicate [01:41] with that data or get information from [01:43] that data using Slack so I can use it [01:46] from my desktop, my mobile, basically [01:48] anywhere I am. And in between those two [01:50] things, I would like a Claude managed [01:52] agent to do the jobs of scraping the [01:55] data, so using tools to get the right [01:57] information, and translating it into way [01:59] that I can understand it. So, to get [02:02] started, I could go to the Claude [02:04] console, go to the new managed agents [02:06] option, and type here in natural [02:08] language to create my agent from [02:10] scratch. Now, this will communicate with [02:12] the Claude API using car commands, and [02:15] will host any necessary code on [02:17] Anthropic's infrastructure. But, I could [02:19] also use the managed agent skill in [02:22] whatever language I prefer. In this [02:23] case, I will pick TypeScript, and this [02:26] skill will use the TypeScript Claude SDK [02:29] to create an agent for me. To do that, [02:32] you'll need to have this version of [02:33] Claude's code or higher, which has the [02:36] built-in managed agent skill. So, in my [02:38] case, I have a version higher than that, [02:40] and can trigger the skill using the [02:42] {slash} Claude API command followed by [02:44] managed agents onboarding. [02:46] So, after I hit enter, it asks me if I [02:48] know what kind of agent I want to build, [02:50] which I don't, but we'll see what it [02:51] does. And it tells me it will walk me [02:53] through these three steps, so tools, [02:56] skills, files, and repos, and then [02:57] environment and identity. Now, [03:00] surprisingly, it does use a lot of [03:01] context, so there might be some [03:03] compaction, but we'll see what happens. [03:05] Anyway, I'll give it a command to create [03:07] a medical agent that reads markdown [03:10] files from a private GitHub repo, [03:12] understands the information like a [03:13] doctor, and lets me communicate with it [03:15] using Slack. And then it recommends me [03:17] to use the pre-built tool set, so read, [03:20] glob, grep, and to not use write, edit, [03:22] bash [03:23] since the doctor has no reason to mutate [03:25] the repo, which makes sense. It also [03:27] asks for the repo URL. So, I'm going to [03:30] tell it to go with its recommended tools [03:31] and permissions, and I'll give it a link [03:33] to the repo. Then after that, it [03:35] suggests what round B and round C should [03:37] be, which since this is a very basic [03:39] agent, they're pretty self-explanatory. [03:41] And then it creates a system prompt for [03:43] my agent, as well as suggest the model I [03:45] should use. Now, I'm going to tell it to [03:46] use Sonnet because I don't want to spend [03:48] too much money on Opus. I'll explain why [03:50] later, but aside from that, this looks [03:52] good, and I'll provide it with these [03:53] credentials as well as the language I [03:55] want. And now it's created the two files [03:58] for me in TypeScript. The first one is [04:00] the setup, which will set up things like [04:02] the environment, the agent, and any [04:04] necessary skills inside Anthropic's [04:06] infrastructure. The second is the [04:08] runtime, which is actually going to be [04:10] the thing that communicates with the [04:12] Anthropic servers and gives that [04:14] information to Slack. So, I'll go ahead [04:16] and set this all up and show you what [04:17] it's like when it's finished. So, after [04:19] running this setup file, it gives me an [04:21] environment ID, which is over here, and [04:24] it also gives me the agent ID, which is [04:26] also over here. Now, as I mentioned [04:28] earlier, these things are created on [04:30] Anthropic's infrastructure. So, inside [04:32] the Claude console, I can see my agent [04:34] over here, as well as the environment [04:36] I've just created. I've also created my [04:38] Slack app and have put all the [04:40] information inside my {dot} env file for [04:42] this app.ts file to use, which means if [04:45] I run that file, it should run my Slack [04:47] bot. So, I can ask it, "What model are [04:49] you using?" and after a while, it [04:51] responds saying, "I'm Claude, made by [04:53] Anthropic. Is there something medical I [04:55] can help you with?" This is very cool. [04:57] But, what's even cooler is that I can [04:59] see the session here in the Claude [05:00] console. Yes, I've been testing this a [05:02] few times. And here we get more details [05:04] of what happened. So, if I close this to [05:05] make it bigger, you can see the question [05:07] the user asked, then it used the Slack [05:10] message tool, and then the agent [05:12] responded. Now, I forgot to mention [05:13] earlier how the pricing works for this. [05:15] So, if we go to the documentation, we [05:17] can see that all tokens used by the [05:19] managed agents are priced with the [05:21] pricing model that the Claude API uses, [05:23] which is over here. So, unfortunately, [05:26] your Pro, Max, or Team subscription [05:28] isn't useful for managed agents. But, as [05:30] well as paying for tokens, you'll also [05:32] have to pay for sessions, which is 8 [05:34] cents per session hour, and this is only [05:37] when the session is running. So, if I go [05:39] back to the Claude console and click on [05:40] sessions, for all of these idle [05:42] sessions, I'm not being charged. Okay, [05:44] let's do something a bit more [05:45] interesting with this bot. I'm going to [05:46] ask it, "Based on what you know about me [05:48] medically, is it okay for me to eat [05:50] calamari?" Now, here it's gone ahead and [05:53] ran the bash tool to get information [05:55] from the repo. It's done two file reads, [05:58] and then a few seconds later, it sent me [06:00] a Slack message, which tells me I should [06:01] be cautious with calamari because I'm [06:04] allergic to shrimp, which is true. I [06:06] would say it's given too much [06:07] information, talking about my itchy [06:09] tongue to swelling throat and so on. [06:11] But, to be honest, it works really well. [06:13] In fact, I did tweak the agent a bit. [06:15] So, if we click here in the Claude [06:17] console, we can see there are three [06:18] versions, indicating that this agent has [06:20] been changed three times. I changed the [06:22] system prompt to make it sound a bit [06:24] more human-like, and changed the model [06:26] from Opus to Sonnet. But, here in the [06:28] UI, I can change the agent system [06:30] prompt, I can change the model, and the [06:32] tools it has access to, which is useful [06:34] for testing. So, that's pretty much it. [06:36] Apart from a few small code tweaks that [06:39] I figured out by going back and forth [06:40] with Claude, that is how easy it is to [06:43] build an agent. No need to learn how the [06:46] Claude agent SDK works. You can just [06:48] communicate with Claude using a skill [06:49] and create your very own agent. But, how [06:52] does all of this work under the hood? [06:54] So, Anthropic have written a very [06:55] detailed article on how everything [06:57] works, which I'll have a link for in the [06:59] description, but I would say the whole [07:02] architecture is made up of three key [07:04] components. So, the session harness and [07:06] orchestration. This is not to say the [07:07] sandbox and tools are less important, [07:09] but I would say these three are very [07:11] unique. So, the harness, also known as [07:13] the hands of the system, is where the [07:15] Claude model is used, and it's also [07:18] known as a stateless router because it [07:19] routes tool calls, resources, and MCPs [07:22] to their relevant place, or runs code [07:24] and edits files in a sandbox [07:27] environment. Now, I'll talk about the [07:28] benefits of having tool calls separate [07:30] from the actual harness itself, but the [07:32] session here is like the memory of the [07:34] system and contains append-only logs of [07:37] the harness. Now, you may be thinking [07:39] the harness is Claude code, but it's [07:41] actually a custom-made harness for the [07:43] managed agents. I'll also explain why a [07:45] bit later on. And finally, the [07:47] orchestrator is what decides what mode [07:50] the harness should be in, so build, [07:52] plan, and so on, and importantly, it [07:55] creates a new harness if this one fails. [07:58] So, imagine you have a harness here that [08:00] fails or or goes down, the orchestrator [08:02] can create a new harness, and because [08:04] the session logs are separate from the [08:06] harness itself, the new harness can read [08:09] the logs to find out what's been going [08:10] on and exactly where to continue from. [08:13] In fact, the whole thing is built for [08:14] scale. So, you can have as many models [08:17] and as many environments as you want, [08:19] and the architecture will be able to [08:20] handle it. Also, another benefit of this [08:22] architecture is the security aspect. So, [08:25] if I go back to the Claude console and [08:27] click on the credentials vault, you can [08:29] see that the credentials are stored in a [08:31] secure location. Now, if I have an agent [08:33] running locally, that could be the {dot} [08:35] env file or something custom I have in [08:36] place, but if I'm using the Claude [08:39] console or the UI, then all the [08:40] credentials are stored here. And the [08:42] beauty of that is these credentials are [08:44] called at runtime. So, if you have an [08:46] MCP-specific API key or tool-specific [08:48] key, then the harness or the model [08:50] doesn't know anything about it, and it [08:52] can't have access to it. So, for [08:54] example, if I needed to call the weather [08:56] MCP tool, and I had that API key, then [08:59] the harness would call the tool, and the [09:01] API key would exist within the tool call [09:03] itself or the MCP, and it would be used [09:06] at runtime. Similarly, if the sandbox [09:09] needed to use a key, then that would be [09:11] stored in the vault, and that would also [09:12] be used at runtime, and the model [09:14] wouldn't know about it. In fact, it [09:16] doesn't even know about its own [09:17] Anthropic API key since that is also [09:20] used at runtime. I highly recommend [09:22] reading the rest of this article to get [09:24] a detailed overview of how the whole [09:25] thing was put together, but it is very [09:28] unique. Honestly, I really enjoy [09:30] creating Claude managed agents. I mean, [09:32] there are people out there who think [09:33] this is going to die just like the Open [09:35] AI agents. If Open AI agents aren't [09:37] dead, please let me know in the comments [09:38] cuz I don't hear much about them, but I [09:40] think this is going to stick around for [09:42] a long time purely because it's very [09:44] easy to create an agent. You don't have [09:46] to learn about SDKs, you don't have to [09:48] use the terminal to create one if you [09:50] don't want to, and you can pretty much [09:53] do all the hooking up to MCP servers, [09:55] tools, all inside the UI, which is very [09:58] nice for people who don't want to write [10:00] any code. But, I do have to say one kind [10:03] of disappointing thing is the fact that [10:04] you have to pay using API tokens or pay [10:07] using API billing because as someone who [10:10] is a Claude Pro subscriber, I would love [10:12] to use my limits, so the limits that are [10:14] within the Pro range on managed agents [10:17] so I don't have to pay for two different [10:19] things. But, in all fairness, it's not [10:21] insanely expensive if you use a cheaper [10:24] model like Sonic or Haiku. And even [10:26] though it's kind of curated, so as you [10:29] can see it gives you access to Notion, [10:31] Slack MCP services, and so on. But, if [10:33] you want to create something that [10:35] doesn't exist within those bounds, then [10:38] you just have to go ahead and write your [10:39] own code, which Open Claude is very good [10:41] for. I mean, Open Claude is super open, [10:43] it's in the name, and has many channels [10:46] from Telegram to Discord to WhatsApp. [10:48] Whereas, if you want to do the same [10:49] thing with managed agents, you may have [10:51] to code up your own solution or get [10:53] Claude to code one up for you. So, the Claude team have released Claude managed agents, which is what they call the next evolution after the agent SDK. This is something that lets you create custom agents without writing a single line of code, and these agents run on Anthropic's own managed infrastructure, which has a very cool architecture that is perfect for shipping secure, scalable, production-ready agents that can do pretty much anything you want, so answer customer queries based on your knowledge base, or even do regular research for you and send it to you at a specific time using Slack. But, why would anyone use this instead of installing Claude's code on their own VPS? Hit subscribe, and let's get into it. But, before we get into it, let's talk a bit about Anthropic because they've been busy these last few weeks. I mean, they recently prevented people from using the Claude subscription on third-party tools like Open Claude, which people think is because of managed agents, but they're not exactly the same thing. I mean, I would say Open Claude is a bit like Linux. It's a tinkerer's agent. You pick your own hardware, you pick your own model, you deal with security and everything in between. Whereas, Claude managed agents is, dare I say it, like Apple because you don't need to do any of those things. Claude takes care of the infrastructure, the security. You just tell it what you want in simple English, and it will go ahead and build your perfect agent. In fact, let me show you how easy it is by going through a simple, yet very personal example. I have a private GitHub repo that contains all my medical information scraped from the NHS app, and I want to communicate with that data or get information from that data using Slack so I can use it from my desktop, my mobile, basically anywhere I am. And in between those two things, I would like a Claude managed agent to do the jobs of scraping the data, so using tools to get the right information, and translating it into way that I can understand it. So, to get started, I could go to the Claude console, go to the new managed agents option, and type here in natural language to create my agent from scratch. Now, this will communicate with the Claude API using car commands, and will host any necessary code on Anthropic's infrastructure. But, I could also use the managed agent skill in whatever language I prefer. In this case, I will pick TypeScript, and this skill will use the TypeScript Claude SDK to create an agent for me. To do that, you'll need to have this version of Claude's code or higher, which has the built-in managed agent skill. So, in my case, I have a version higher than that, and can trigger the skill using the {slash} Claude API command followed by managed agents onboarding. So, after I hit enter, it asks me if I know what kind of agent I want to build, which I don't, but we'll see what it does. And it tells me it will walk me through these three steps, so tools, skills, files, and repos, and then environment and identity. Now, surprisingly, it does use a lot of context, so there might be some compaction, but we'll see what happens. Anyway, I'll give it a command to create a medical agent that reads markdown files from a private GitHub repo, understands the information like a doctor, and lets me communicate with it using Slack. And then it recommends me to use the pre-built tool set, so read, glob, grep, and to not use write, edit, bash since the doctor has no reason to mutate the repo, which makes sense. It also asks for the repo URL. So, I'm going to tell it to go with its recommended tools and permissions, and I'll give it a link to the repo. Then after that, it suggests what round B and round C should be, which since this is a very basic agent, they're pretty self-explanatory. And then it creates a system prompt for my agent, as well as suggest the model I should use. Now, I'm going to tell it to use Sonnet because I don't want to spend too much money on Opus. I'll explain why later, but aside from that, this looks good, and I'll provide it with these credentials as well as the language I want. And now it's created the two files for me in TypeScript. The first one is the setup, which will set up things like the environment, the agent, and any necessary skills inside Anthropic's infrastructure. The second is the runtime, which is actually going to be the thing that communicates with the Anthropic servers and gives that information to Slack. So, I'll go ahead and set this all up and show you what it's like when it's finished. So, after running this setup file, it gives me an environment ID, which is over here, and it also gives me the agent ID, which is also over here. Now, as I mentioned earlier, these things are created on Anthropic's infrastructure. So, inside the Claude console, I can see my agent over here, as well as the environment I've just created. I've also created my Slack app and have put all the information inside my {dot} env file for this app.ts file to use, which means if I run that file, it should run my Slack bot. So, I can ask it, "What model are you using?" and after a while, it responds saying, "I'm Claude, made by Anthropic. Is there something medical I can help you with?" This is very cool. But, what's even cooler is that I can see the session here in the Claude console. Yes, I've been testing this a few times. And here we get more details of what happened. So, if I close this to make it bigger, you can see the question the user asked, then it used the Slack message tool, and then the agent responded. Now, I forgot to mention earlier how the pricing works for this. So, if we go to the documentation, we can see that all tokens used by the managed agents are priced with the pricing model that the Claude API uses, which is over here. So, unfortunately, your Pro, Max, or Team subscription isn't useful for managed agents. But, as well as paying for tokens, you'll also have to pay for sessions, which is 8 cents per session hour, and this is only when the session is running. So, if I go back to the Claude console and click on sessions, for all of these idle sessions, I'm not being charged. Okay, let's do something a bit more interesting with this bot. I'm going to ask it, "Based on what you know about me medically, is it okay for me to eat calamari?" Now, here it's gone ahead and ran the bash tool to get information from the repo. It's done two file reads, and then a few seconds later, it sent me a Slack message, which tells me I should be cautious with calamari because I'm allergic to shrimp, which is true. I would say it's given too much information, talking about my itchy tongue to swelling throat and so on. But, to be honest, it works really well. In fact, I did tweak the agent a bit. So, if we click here in the Claude console, we can see there are three versions, indicating that this agent has been changed three times. I changed the system prompt to make it sound a bit more human-like, and changed the model from Opus to Sonnet. But, here in the UI, I can change the agent system prompt, I can change the model, and the tools it has access to, which is useful for testing. So, that's pretty much it. Apart from a few small code tweaks that I figured out by going back and forth with Claude, that is how easy it is to build an agent. No need to learn how the Claude agent SDK works. You can just communicate with Claude using a skill and create your very own agent. But, how does all of this work under the hood? So, Anthropic have written a very detailed article on how everything works, which I'll have a link for in the description, but I would say the whole architecture is made up of three key components. So, the session harness and orchestration. This is not to say the sandbox and tools are less important, but I would say these three are very unique. So, the harness, also known as the hands of the system, is where the Claude model is used, and it's also known as a stateless router because it routes tool calls, resources, and MCPs to their relevant place, or runs code and edits files in a sandbox environment. Now, I'll talk about the benefits of having tool calls separate from the actual harness itself, but the session here is like the memory of the system and contains append-only logs of the harness. Now, you may be thinking the harness is Claude code, but it's actually a custom-made harness for the managed agents. I'll also explain why a bit later on. And finally, the orchestrator is what decides what mode the harness should be in, so build, plan, and so on, and importantly, it creates a new harness if this one fails. So, imagine you have a harness here that fails or or goes down, the orchestrator can create a new harness, and because the session logs are separate from the harness itself, the new harness can read the logs to find out what's been going on and exactly where to continue from. In fact, the whole thing is built for scale. So, you can have as many models and as many environments as you want, and the architecture will be able to handle it. Also, another benefit of this architecture is the security aspect. So, if I go back to the Claude console and click on the credentials vault, you can see that the credentials are stored in a secure location. Now, if I have an agent running locally, that could be the {dot} env file or something custom I have in place, but if I'm using the Claude console or the UI, then all the credentials are stored here. And the beauty of that is these credentials are called at runtime. So, if you have an MCP-specific API key or tool-specific key, then the harness or the model doesn't know anything about it, and it can't have access to it. So, for example, if I needed to call the weather MCP tool, and I had that API key, then the harness would call the tool, and the API key would exist within the tool call itself or the MCP, and it would be used at runtime. Similarly, if the sandbox needed to use a key, then that would be stored in the vault, and that would also be used at runtime, and the model wouldn't know about it. In fact, it doesn't even know about its own Anthropic API key since that is also used at runtime. I highly recommend reading the rest of this article to get a detailed overview of how the whole thing was put together, but it is very unique. Honestly, I really enjoy creating Claude managed agents. I mean, there are people out there who think this is going to die just like the Open AI agents. If Open AI agents aren't dead, please let me know in the comments cuz I don't hear much about them, but I think this is going to stick around for a long time purely because it's very easy to create an agent. You don't have to learn about SDKs, you don't have to use the terminal to create one if you don't want to, and you can pretty much do all the hooking up to MCP servers, tools, all inside the UI, which is very nice for people who don't want to write any code. But, I do have to say one kind of disappointing thing is the fact that you have to pay using API tokens or pay using API billing because as someone who is a Claude Pro subscriber, I would love to use my limits, so the limits that are within the Pro range on managed agents so I don't have to pay for two different things. But, in all fairness, it's not insanely expensive if you use a cheaper model like Sonic or Haiku. And even though it's kind of curated, so as you can see it gives you access to Notion, Slack MCP services, and so on. But, if you want to create something that doesn't exist within those bounds, then you just have to go ahead and write your own code, which Open Claude is very good for. I mean, Open Claude is super open, it's in the name, and has many channels from Telegram to Discord to WhatsApp. Whereas, if you want to do the same thing with managed agents, you may have to code up your own solution or get Claude to code one up for you., Anthropic's Managed Agents Are Different (Here's Why), YouTube, 01kp1wtw517t2fkkmczn35ntah.