How Harness-as-a-Service Will Change Agents

How Harness-as-a-Service Will Change Agents 🔗 **Source:** [How Harness-as-a-Service Will Change Agents](https://www.youtube.com/watch?v=jvqQ8VlhO-w) > [!note]+ Document Notes > tags: #artificial-intelligence #agents #infrastructure #cloud-computing #enterprise-software > author: AI Daily Brief > source: YouTube > reference: https://www.youtube.com/watch?v=jvqQ8VlhO-w > date created: 2025-01-30 14:30 > > # 🏷️ Title: How Harness-as-a-Service Will Change Agents >  ## 🧰 Action Inbox - [H1] Evaluate Cursor SDK for agent development — new platform enables local/cloud agent building #artificial-intelligence - [H2] Research harness-as-a-service market opportunities — emerging infrastructure category #infrastructure - [H3] Assess big tech AI earnings impact on enterprise strategy — massive growth in cloud AI services #enterprise-software ## 🧭 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 - [Cursor SDK](https://cursor.com/sdk) - [AI Daily Brief](https://aidailybrief.ai) ## 📝 Summary - **What happened**: Major tech companies reported massive AI earnings growth, while Cursor launched their SDK introducing "harness-as-a-service" - pre-built agent runtime infrastructure - **Why it matters**: This represents a shift from building agents from scratch to using managed infrastructure, similar to how cloud computing democratized application development - **Key stakeholders**: AI developers, enterprise customers, cloud providers, agent builders - **Risks and tradeoffs**: Potential vendor lock-in vs. faster development; standardization vs. customization flexibility - **Bottom line**: AI infrastructure is maturing rapidly with both demand (evidenced by earnings) and tooling (harness-as-a-service) reaching new levels ## 📖 Main Content The episode covers two major developments: big tech earnings showing explosive AI growth and the emergence of "harness-as-a-service" as a new infrastructure category. **Big Tech AI Earnings Highlights:** - Google Cloud: 63% year-over-year growth, $460B backlog - AWS: 28% growth, fastest in 4 years - Azure: 39% growth with steady acceleration - Meta: 33% revenue growth despite user metric decline These numbers demonstrate that AI demand is no longer speculative but driving real business growth across all major cloud providers. **Harness-as-a-Service Concept:** The host argues we're seeing the emergence of a new infrastructure category where companies sell access to pre-built agent runtimes. This includes the "harness" - the environment and tools that turn LLMs into functional agents capable of executing tasks. The evolution is described in three phases: 1. **Weights Phase**: Focus on bigger, better models 2. **Context Phase**: Optimizing what models see (prompts, RAG) 3. **Harness Engineering Phase**: Creating intelligent environments for models to operate in ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/jvqQ8VlhO-w?t=0) — [H1] Big Tech AI Earnings Overview - Key idea: Major tech companies show unprecedented AI-driven growth - Evidence: "Google Cloud plus 63% year-over-year, Microsoft Azure plus 40%, Meta revenue plus 33%, AWS plus 28%" #### ⏱️ [05:30](https://youtu.be/jvqQ8VlhO-w?t=330) — [H1] Google's AI Dominance - Key idea: Google delivered massive beats across AI-related businesses - Evidence: "Google Cloud has experienced 63% revenue growth... $460 billion backlog in new orders, up from 240 at the end of Q4" #### ⏱️ [08:45](https://youtu.be/jvqQ8VlhO-w?t=525) — [H2] Search Revenue Surprise - Key idea: AI enhanced rather than cannibalized Google's core search business - Evidence: "The prevailing narrative was that AI would cannibalize Google's core business... The opposite is happening. Search ad revenue grew 19%" #### ⏱️ [12:20](https://youtu.be/jvqQ8VlhO-w?t=740) — [H2] Amazon's Infrastructure Play - Key idea: Amazon positioned as picks-and-shovels infrastructure provider - Evidence: "Amazon said they had added more server capacity than any other company in 2025 and plan to accelerate construction even more this year" #### ⏱️ [15:10](https://youtu.be/jvqQ8VlhO-w?t=910) — [H2] Microsoft's Middle Position - Key idea: Microsoft caught between aggressive competitors with mixed results - Evidence: "Microsoft is starting to fall behind their peers who are making bigger moves" #### ⏱️ [18:00](https://youtu.be/jvqQ8VlhO-w?t=1080) — [H3] Meta's CapEx Concerns - Key idea: Meta's aggressive spending on AI infrastructure concerns investors - Evidence: "Meta was the biggest loser from big tech earnings... The market still hates their CapEx spend" #### ⏱️ [22:30](https://youtu.be/jvqQ8VlhO-w?t=1350) — [H1] Harness-as-a-Service Introduction - Key idea: New infrastructure category emerging for agent development - Evidence: "Cursor SDK handles the harness, sandboxing, computer use, demo videos, GitHub integration, etc." #### ⏱️ [25:00](https://youtu.be/jvqQ8VlhO-w?t=1500) — [H1] Three Phases of Agent Evolution - Key idea: Agent development evolved from weights to context to harness engineering - Evidence: "The biggest shift in AI agents had nothing to do with making models smarter, but was instead about making the environment around them smarter" #### ⏱️ [28:15](https://youtu.be/jvqQ8VlhO-w?t=1695) — [H2] Sam Altman on Harness Importance - Key idea: Industry leaders recognize harness as critical as models themselves - Evidence: "Hard to overstate how critical it is. I no longer think of the harness and the model as these entirely separable things" #### ⏱️ [32:00](https://youtu.be/jvqQ8VlhO-w?t=1920) — [H2] Open Claw Era Limitations - Key idea: Current open-source approach requires extensive technical setup - Evidence: "You had to do everything from picking the model to write the system prompt to define the tool to wire the agent loop" #### ⏱️ [35:45](https://youtu.be/jvqQ8VlhO-w?t=2145) — [H2] PC Era Analogy - Key idea: Harness-as-a-service parallels transition from computer kits to pre-built PCs - Evidence: "The productivity revolution of the 1990s happened because users got Dell desktops, not because more people learn to assemble motherboards" #### ⏱️ [38:20](https://youtu.be/jvqQ8VlhO-w?t=2300) — [H2] Gmail Agent Demo - Key idea: Real-world example of cursor SDK enabling agent integration - Evidence: "Jack Driscoll showed off a demo... cursor agent embedded directly into Gmail" #### ⏱️ [41:30](https://youtu.be/jvqQ8VlhO-w?t=2490) — [H3] Performance Benchmarks - Key idea: Harness choice significantly impacts model performance - Evidence: "GPT-5.5 went from 61.5% to 87.2% by switching to cursor... same model, same week, two harnesses, two different functional results" #### ⏱️ [44:00](https://youtu.be/jvqQ8VlhO-w?t=2640) — [H3] Early SDK Applications - Key idea: Developers quickly building innovative applications with new platform - Evidence: "Tejas Vavery built a bug-catching agent that can work on his production code base and can see how the app is performing" ## 📚 References - [Cursor SDK GitHub](https://github.com/cursor/sdk) - [Google Earnings Report](https://abc.xyz/investor/) - [Amazon Q1 Results](https://ir.aboutamazon.com/) - [Microsoft Earnings](https://www.microsoft.com/en-us/investor/) - [Meta Financial Results](https://investor.fb.com/) ## 🗣️ Paraphrase - Big tech earnings demonstrate that AI demand has moved from speculation to driving real revenue growth across all major cloud providers - A new infrastructure category called "harness-as-a-service" is emerging, providing pre-built agent runtimes instead of requiring developers to build everything from scratch - This shift parallels the PC revolution - moving from hobbyist computer kits to plug-and-play systems that democratized computing access ## 📰 Newsletter **AI Infrastructure Reaches New Maturity** This week brought explosive AI earnings from big tech and the launch of Cursor's SDK, signaling a new phase in AI infrastructure maturity. Google Cloud's 63% growth and Amazon's $200B infrastructure investment show that AI demand is driving real business results. Meanwhile, Cursor's SDK introduces "harness-as-a-service" - pre-built agent runtimes that handle sandboxing, tool integration, and execution environments. This could democratize agent development the same way cloud computing democratized application development. **Key Benefits:** - Faster agent development with pre-built infrastructure - Focus on business logic instead of technical plumbing - Better performance through optimized harnesses **Getting Started:** - Explore the Cursor SDK for agent development - Consider how managed agent services could accelerate your projects - Evaluate harness-as-a-service options for enterprise use cases Resources: [Cursor SDK](https://cursor.com/sdk), [AI Daily Brief](https://aidailybrief.ai), [Agent OS Program](https://aidailybrief.ai/agent-os) ## ✅ Action Items #### DevRel - Developer advocacy: Create workshops on harness-as-a-service concepts and best practices - Education: Develop curriculum comparing DIY agent development vs. managed harness services - Developer experience: Evaluate SDK documentation and onboarding flows for agent platforms - Developer marketing: Position harness-as-a-service as democratizing force in AI development - Community management: Foster discussions around agent architecture patterns and tooling choices #### OSPO - Policy & Governance: Establish guidelines for evaluating proprietary vs. open-source agent harnesses - License Compliance: Review licensing implications of managed agent services and SDKs - Security Management: Assess security posture of harness-as-a-service platforms - Software Inventory: Track agent runtime dependencies and harness provider relationships - Developer Support: Provide training on agent architecture decisions and vendor evaluation - Strategic Alignment: Align agent development strategy with broader infrastructure modernization #### Community - Maintainers: Support development of open-source harness alternatives and standards - Contributors: Create pathways for contributing to agent runtime and harness projects - Sponsors: Evaluate funding opportunities for critical agent infrastructure projects - Ecosystem: Foster collaboration between harness providers and application developers ## 🧪 Hypotheses - Harness-as-a-service will follow similar adoption patterns to cloud computing, starting with startups and expanding to enterprises - Performance differences between harnesses will become a key differentiator, similar to database or compute engine choices - Open-source harness alternatives will emerge to compete with proprietary platforms like Cursor SDK - The quality of the harness will matter more than marginal model improvements for most agent applications ### 📊 Methodology - Benchmark agent performance across different harness platforms using standardized coding and reasoning tasks - Survey developer productivity gains when migrating from custom harness implementations to managed services - Analyze cost-benefit tradeoffs between building custom agent infrastructure vs. using harness-as-a-service - Track enterprise adoption patterns of managed vs. self-hosted agent runtimes ### ❓ Questions - How will harness standardization affect agent portability across platforms? - What happens to innovation when critical agent infrastructure becomes commoditized? - Will harness-as-a-service create new vendor lock-in risks for AI applications? ### 🔮 Future Research - Comparative analysis of emerging harness-as-a-service platforms and their architectural choices - Investigation of open-source harness development and standardization efforts - Study of enterprise decision-making factors in agent infrastructure procurement ### 💡 Implications - Practice: Developers can focus on business logic rather than infrastructure when building agents - Policy: Organizations need evaluation frameworks for agent platform vendor selection - Security: Centralized harness services create new attack surfaces and dependency risks - Education: Developer training must cover agent architecture patterns and platform evaluation - Funding: Investment will likely flow toward both proprietary platforms and open-source alternatives ### 🧠 Learning Styles Assessment Reading level estimate: 12th Grade #### 👀 Visual learners - Present: Performance benchmark charts and earnings growth comparisons - Improve: Architecture diagrams showing harness components, evolution timeline graphics #### 🔊 Aural learners - Present: Full podcast episode with expert commentary and real-world examples - Improve: Interview segments with developers using new platforms #### 🗣️ Verbal learners - Present: Clear analogies (PC era comparison), defined technical terms, concrete examples - Improve: Glossary of agent development terminology, more industry metaphors #### 🧩 Kinesthetic learners - Present: Code examples and demo descriptions of working applications - Improve: Hands-on tutorials, step-by-step SDK setup guides, interactive demos #### 🧠 Logical learners - Present: Three-phase evolution framework, earnings data tables, performance benchmarks - Improve: Decision matrices for platform selection, cost-benefit analysis frameworks #### 🧑‍🤝‍🧑 Social learners - Present: Developer community examples, industry expert quotes - Improve: Discussion forums for platform comparison, developer community showcases #### 🧘‍♀️ Solitary learners - Present: Detailed technical explanations, self-contained examples - Improve: Self-assessment frameworks for platform fit, independent research guides #### 🌿 Naturalistic learners - Present: Technology ecosystem evolution patterns - Improve: Connections to broader infrastructure evolution cycles, sustainability considerations ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure, descriptive timestamps - Improve: Better alt text for any embedded charts or graphics #### Keyboard navigation - Present: Standard video player controls, logical tab order - Improve: Skip-to-content links, keyboard shortcuts for timestamp navigation #### Contrast and readability - Present: Standard text formatting with good contrast - Improve: High-contrast mode support, customizable text sizing #### Text resizing and reflow - Present: Responsive text layout - Improve: Ensure no content clipping at 200% zoom #### Media and captions - Present: Full transcript provided - Improve: Synchronized captions with timestamp markers, audio descriptions for charts #### Forms and interactions - Present: Basic link interactions - Improve: Clear focus indicators, accessible timestamp navigation #### Standards reference - WCAG 2.2 AA compliance needed for video player controls and transcript formatting ### 🌍 DEIB Assessment #### Cultural breadth - Present: US-focused tech industry perspective - Improve: Global AI development perspectives, emerging market considerations #### Inclusive language - Present: Technical terms clearly defined, neutral professional tone - Improve: Avoid assuming familiarity with Silicon Valley culture #### Neurodiversity support - Present: Clear structure with headers and timestamps - Improve: Content warnings for rapid topic changes, more frequent summaries #### Representation - Present: Mix of company perspectives and individual developer examples - Improve: Include voices from underrepresented groups in tech, diverse geographic representation #### Psychological safety - Present: Professional, informative tone without judgment - Improve: Acknowledge learning curve challenges, normalize not knowing technical concepts ### ✅ Logical Validity #### Strong reasoning - Deductive strength: Clear logical progression from individual company earnings --- ## Full Transcript [00:00] Today on the AI Daily Brief, the [00:02] emergence of harness as a service, what [00:04] it means for the agentic era. [00:07] And before that in the headlines, [00:09] a big tech AI earnings blowout. [00:11] The AI Daily Brief is a daily podcast [00:13] and video about the most important news [00:15] and discussions in AI. [00:18] All right, friends, quick announcements [00:19] before we dive in. If you haven't yet, [00:21] go check out the new Agent OS program. [00:24] It's a tool-agnostic adaptable system [00:26] for building an agentic operating [00:27] system, and I have a feeling after you [00:29] listen to today's harness as a service [00:31] episode that you will want to dig in [00:32] even more. You can find that off of the [00:34] main site, aidailybrief.ai. [00:37] Today is one of those rare days where [00:39] the headlines are all around the same [00:41] theme, and that theme is, of course, big [00:45] tech earnings. [00:46] And to not bury the lead, let's go over [00:48] to Sheharyar Khan who writes, [00:49] "Hard to take the AI bubble argument [00:51] seriously when some of the largest [00:53] companies on Earth are still putting up [00:54] these growth numbers. Google Cloud plus [00:57] 63% year-over-year. Microsoft Azure plus [01:00] 40% year-over-year. Meta revenue plus [01:03] 33% year-over-year. AWS plus 28% [01:07] year-over-year." [01:09] We're going to go through all of these [01:10] and talk about the winners and losers [01:11] and what it means for the market's [01:12] assessment of AI overall. Google was the [01:15] clear winner on big tech earnings night, [01:17] delivering huge beats across the board. [01:19] They reported 22% top-line revenue [01:21] growth, but as I just mentioned, the big [01:23] numbers were in their AI-related [01:24] businesses. Google Cloud has experienced [01:27] 63% revenue growth over the past year. [01:30] They also reported a $460 billion [01:32] backlog in new orders, up from 240 at [01:35] the end of Q4. Their new deal with [01:37] Anthropic contributes a decent chunk of [01:39] that growth, but it still shows that GPU [01:41] demand is off the charts. Analyst Joseph [01:43] Carlson posted a chart of Google's cloud [01:45] backlog going exponential and commented, [01:47] "This is so crazy it literally looks [01:49] fake." Gemini growth was similarly [01:51] strong. Google reported a 40% surge in [01:54] paid enterprise customers [01:55] quarter-over-quarter, meaning maybe I'm [01:56] going to have to eat my hat rating [01:58] Google as low as I did on enterprise in [02:00] the AI Lab Power Rankings. [02:02] In addition, Google's infrastructure is [02:03] now processing 16 billion tokens a [02:05] minute, up 60% quarter-over-quarter. [02:08] Even search, which is tangentially [02:09] related to some of Google's AI changes, [02:11] is experiencing a boom with search [02:13] revenue up 19% year-over-year. Google is [02:15] also maintaining very healthy profit [02:17] margins, hitting 62.6 billion in net [02:19] income for an 81% year-over-year gain. [02:22] CEO Sundar Pichai told analysts that AI [02:24] is now the largest tailwind for cloud, [02:26] commenting, "Our enterprise AI solutions [02:28] have become our primary growth driver [02:29] for cloud for the first time in Q1." He [02:32] added, though, "We are [02:33] compute-constrained in the near term. [02:35] Our cloud revenue would have been higher [02:36] if we were able to meet the demand." [02:39] CapEx targets, meanwhile, remained [02:41] pretty anchored. Google slid up their [02:43] forecast slightly, increasing this [02:44] year's guide from a range of 175 to 185 [02:47] billion to a range of 180 to 190 [02:49] billion, although right now they're not [02:51] coming close to hitting that target. [02:53] They only recorded 35.7 billion in CapEx [02:55] spending for Q1, which annualizes out to [02:57] 140 billion and change. Now, Google [03:00] might be padding the back end of the [03:01] year in expectation of rising costs, but [03:03] because this came in slightly below [03:04] analyst forecast, the market interpreted [03:06] it as capital discipline. Google was up [03:08] 7% in overnight trading. [03:11] Chubby, meanwhile, argued that Google [03:12] has cracked the AI monetization thesis. [03:15] While the absurdly high cloud growth was [03:16] the loudest signal, Chubby added, "The [03:18] real story is search. The prevailing [03:20] narrative was that AI would cannibalize [03:22] Google's core business. People get [03:23] answers from chatbots, stop Googling. [03:25] The opposite is happening. Search ad [03:27] revenue grew 19%. Queries hit an [03:29] all-time high. Google turned the biggest [03:31] existential threat to its business into [03:32] a growth accelerator." [03:34] Amazon's earnings were also extremely [03:36] solid, suggesting a return to form on [03:38] cloud. Top-line revenue growth came in [03:40] at 17% year-over-year with net profit up [03:42] 77%. Amazon attributed the growth in [03:45] profit to pre-tax income from their [03:46] Anthropic investment, so this number [03:48] could say more about Anthropic than it [03:50] does about Amazon. AWS revenue was up [03:52] 28% compared to last year with Amazon's [03:55] dominant cloud unit coming out of a [03:56] major slowdown after achieving a low of [03:58] 12% growth in 2023. Growth has been [04:00] accelerating every quarter since and is [04:02] now back at a healthy level. The number [04:04] beat analyst forecast and would have [04:05] been viewed very solid if not for the [04:07] direct comparison to Google Cloud's [04:08] incredible 63%. [04:10] To the extent that there was an issue, [04:11] it was in Amazon's costly AI buildout. [04:14] Amazon said they had added more server [04:16] capacity than any other company in 2025 [04:18] and plan to accelerate construction even [04:20] more this year. Their capital spending [04:22] for Q1 was 43.2 billion, putting them [04:24] slightly behind their $200 target pace [04:26] for the full year, but that remains a [04:28] 60% jump from last year and the largest [04:30] overall number between the four big tech [04:32] hyperscalers. Amazon is now essentially [04:34] spending every dollar they make on the [04:36] AI buildout, driving free cash flows for [04:38] the quarter up from almost 26 billion [04:40] this time last year down to 1.2 billion [04:42] this quarter. [04:43] CEO Andy Jassy dismissed concerns, [04:45] stating the company would keep spending [04:47] to keep up with demand. He added that he [04:49] was confident that additional spending [04:50] would translate into strong profits [04:52] because most of the new supplies are [04:53] already spoken for. And, of course, with [04:55] OpenAI now partnered with AWS alongside [04:57] Anthropic, that is probably a fairly [04:59] safe bet. The Wall Street Journal called [05:01] the AWS expansion prescient, writing, [05:03] "The growing demand for chatbots and [05:04] other AI-powered tools is outpacing the [05:06] supply of chips and storage, causing [05:08] outages and surging prices." Jassy also [05:11] noted that Amazon's in-house chips have [05:12] been well received, commenting, "We have [05:14] such demand right now for Trainium from [05:16] various companies who will consume as [05:17] much as we make. I expect over time [05:19] there's a good chance we're going to [05:20] sell racks over the coming years. We [05:22] have to decide how much we're going to [05:23] allocate to the existing demand and how [05:25] much we're going to save to sell as [05:26] racks." He commented that if their chips [05:28] were a standalone business and actually [05:30] booking revenue from AWS, they would be [05:32] sitting at 50 billion in ARR. Jassy [05:34] boasted, "As best as we can tell, our [05:36] custom silicon business is now one of [05:38] the top three data center chip [05:39] businesses in the world, and the speed [05:41] at which we've gotten here is [05:42] extraordinary." [05:44] Now, one of the big themes over the past [05:45] year has been Amazon reasserting their [05:47] dominance as the largest cloud provider. [05:49] There's been far less emphasis on [05:50] training their own models and releasing [05:51] AI products. Instead, they're forming [05:54] strong strategic partnerships and [05:55] putting themselves in prime position as [05:56] a picks and shovels AI infrastructure [05:58] play. Jassy reinforced that strategic [06:00] pivot during the earnings call, [06:01] commenting, "We're in the middle of some [06:03] of the biggest inflections of our [06:04] lifetime. We're well-positioned to lead, [06:06] and I'm very optimistic about what's [06:07] ahead for our customers and for Amazon." [06:09] Still, the market had a bit of a [06:11] lukewarm response. The stock was down [06:13] initially, but ended the overnight [06:14] session up 2.6%. Many attributed this, [06:16] though, to the market already boosting [06:18] Amazon on news of the OpenAI deal [06:20] earlier in the week. The logic is that [06:21] with the stock already elevated, it [06:23] would have taken a big upside surprise [06:24] to push it even further. Most of the [06:26] commentary centered around AWS righting [06:28] the ship with a little help from the [06:29] gigantic compute crunch. Sheharyar again [06:31] wrote, "AWS is now a $152 billion ARR [06:35] business growing 28% per year. That [06:37] marks its fastest growth in nearly four [06:39] years." Wrote Signal, "People [06:41] underestimate how big of a deal it is [06:42] that OpenAI models are now on Bedrock. [06:44] I've met so many companies that [06:46] defaulted to Anthropic and Claude [06:47] because they were already on Bedrock, [06:49] and for a long time that was basically [06:51] the path of least resistance. This is [06:52] huge for OpenAI model accessibility." [06:55] Moving on to Microsoft, the company has [06:57] been sort of in this middle space of the [06:58] three big cloud giants. They didn't [07:00] commit to aggressive growth like Google, [07:02] instead opting for a more conservative [07:04] AI CapEx strategy. They also have scale [07:06] and incumbency disadvantage when [07:07] compared to AWS. Most of what they did [07:10] have was exclusive distribution rights [07:11] to OpenAI's models, and that's now gone. [07:13] At the same time, they're one of the few [07:15] companies selling tokens during a token [07:16] drought, and that counts for a whole [07:18] lot. The big story from earnings was 39% [07:21] growth for Azure. That was right in line [07:22] with expectations and just a one [07:24] percentage point acceleration from the [07:25] previous quarter. Top-line revenue was a [07:27] slight beat, coming in at 82.9 billion [07:29] for an 18% year-over-year gain. In terms [07:31] of forecast, CFO Amy Hood projected this [07:33] growth rate to continue at a steady [07:35] pace, anticipating Azure's 40% growth [07:37] rate to continue for Q2. CapEx guidance [07:39] was lifted by 25 billion to reach 190 [07:42] billion for the year. Hood attributed [07:43] the entire increase to higher component [07:45] prices rather than adding more data [07:46] center projects. On Copilot, CEO Satya [07:49] Nadella said that Microsoft now has 20 [07:51] million paid seats for the enterprise [07:52] add-on, up from 15 million in January. [07:55] Microsoft expects that growth to [07:56] continue into the third quarter. It [07:58] appears that utilization is also up with [08:00] Nadella commenting, "Weekly engagement [08:02] is now at the same level as Outlook as [08:03] more and more users make Copilot a [08:05] habit." Responding to questions about [08:07] losing exclusive access to OpenAI's [08:08] models, Nadella downplayed concerns, [08:10] stating, "We have a frontier model [08:12] royalty-free with all the IP rights that [08:14] we will have access to all the way to [08:16] 32, and we fully plan to exploit it." [08:18] Over the past few months, the market has [08:20] lumped Microsoft in with software stocks [08:21] and included it in the SaaS apocalypse [08:23] with the stock down more than 10% so far [08:25] this year. Unfortunately for them, last [08:27] night's earnings didn't seem to do [08:28] anything to change that viewpoint with [08:30] the stock whipsawing up and down [08:31] throughout the overnight session, but [08:32] ending the night basically flat. [08:34] Ultimately, nothing disastrous or [08:36] spectacular came out of the earnings [08:37] call. They continue to add GPUs to their [08:39] fleet at a tremendous pace, but a pace [08:41] that feels conservative compared to [08:42] Google's numbers. It's a big positive to [08:44] see Copilot growing, but the negative [08:46] take is that 20 million is still a drop [08:47] in the ocean compared to the roughly 320 [08:49] million paid seats for Office 365. [08:52] After earnings were released, but prior [08:53] to the investor call, Gene Munster of [08:55] Deepwater Management wrote, "Microsoft [08:57] is in a tight spot because they can't [08:59] shake the negative narrative. Stock is [09:01] down 1% on solid numbers. This past [09:03] month, Microsoft is up 17% in line with [09:05] the Nasdaq. My take, they need to make a [09:07] statement on the call tonight and make [09:09] it easy for investors to believe that [09:10] they have AI-powered products beyond [09:12] Azure that customers must have. Copilot [09:14] has been a miss." Nadella ultimately [09:16] didn't deliver that big statement, and [09:18] so the company continues to perform like [09:19] a perfectly average tech stock. [09:21] Not a bad place to be during the AI [09:23] boom, but Microsoft is starting to fall [09:25] behind their peers who are making bigger [09:27] moves. [09:28] Now, on the Meta front, the past year [09:30] has been all about rapidly expanding [09:31] CapEx. Beginning in 2024, Mark [09:34] Zuckerberg seemed to make the decision [09:35] that he would continue accelerating data [09:36] center spending regardless of investor [09:38] concerns. Meta now find themselves with [09:40] the largest market cap-adjusted CapEx [09:42] commitments among the tech giants. Over [09:43] the past few quarters, the narrative was [09:45] that revenue growth was accelerating [09:47] fast enough to justify the AI spend, and [09:49] that Zuckerberg had earned the right to [09:50] spend big on data centers by delivering [09:52] solid returns. Analysts were linking [09:54] increasing advertising revenue directly [09:56] to AI optimizations in Meta's ad [09:57] platform. Heading into Wednesday night's [09:59] earnings, the balance between revenue [10:01] growth and CapEx remained the central [10:02] question. The financial reports showed [10:04] that Meta had delivered another record [10:06] quarter. [10:07] Quarterly revenue came in at 56.3 [10:09] billion, up 33% over the past year. Both [10:11] total revenue and net income beat [10:13] analyst forecast by a fairly significant [10:15] margin. On the other side of the ledger, [10:17] Meta hiked capital expenses once again. [10:19] They increased this year's forecast from [10:21] 135 billion to 145 billion. Notably, [10:24] this wasn't because Meta was tacking on [10:26] another data center or two. The CapEx [10:28] hike was instead attributed to [10:29] expectations for higher component [10:31] pricing and additional data center [10:32] costs. Meta also disclosed a [10:34] quarter-over-quarter decrease in daily [10:36] active people, which is Meta's in-house [10:38] metric for user activity. [10:40] The decline was small and Meta blamed [10:41] internet disruptions in Iran alongside a [10:43] restriction on WhatsApp in Russia. [10:45] Still, it's the first time that Meta's [10:46] user metrics have decreased since they [10:48] first began reporting it in 2019. [10:50] Advocating for his data center buildout, [10:52] Mark Zuckerberg told investors, "Every [10:54] sign that we're seeing in our own work [10:55] and across the industry gives us [10:56] confidence in this investment. That [10:58] said, we are very focused on increasing [11:00] the efficiency of our investments." Meta [11:02] CFO Susan Li noted that Meta is [11:03] suffering from the same compute [11:04] shortages everyone else, even though [11:06] most of their AI computers deployed [11:07] towards internal and infrastructure use [11:09] rather than customer use. Li said, "Our [11:12] experience so far has been that we have [11:13] underestimated our compute needs, even [11:15] as we have been ramping capacity [11:17] significantly. So, our expectation is [11:19] that compute will become even more [11:20] central to the business going forward." [11:22] In any other year, this would have been [11:24] a blowout earnings result. A 33% revenue [11:26] growth rate is the highest they've [11:27] achieved since 2021 and up substantially [11:30] from 23.8% for the previous quarter. And [11:32] yet, Meta was the biggest loser from the [11:34] big tech earnings. The market still [11:36] hates their CapEx spend and sent the [11:37] stock plummeting by 5% in overnight [11:39] trading. [11:40] Jim Cramer summed up the market opinion [11:42] when he said, "Meta did not offer enough [11:44] reasons to spend the way that other [11:45] companies did. They just told us that [11:47] they could do better with it. Hence the [11:48] decline tonight." [11:50] If there is one big takeaway from tech [11:52] earnings, it's that the AI boom is in [11:54] full effect. Everything from memory chip [11:56] foundries to data center construction is [11:58] running at 100% in a frankly vain [12:00] attempt to keep up with the endless [12:02] demand for tokens. This is categorically [12:04] different to last year, when many [12:06] analysts still had questions and AI [12:08] optimists were betting on a forthcoming [12:09] boom. That boom is now clearly here and [12:12] it's showing up in big tech's growth [12:13] numbers. There are still, of course, [12:15] reasonable questions around business [12:16] models and profitability, as well as who [12:18] the biggest winners will be. But, AI [12:20] demand is unquestionable at this point [12:22] and appears to still be accelerating. [12:25] For now, that that's going to do it for [12:26] today's headlines. Next up, the main [12:28] episode. [12:29] Welcome back to the AI Daily Brief. [12:31] Nominally, our topic today is Cursor's [12:34] new Cursor SDK. It's a platform where, [12:37] as Cursor's Li Robinson put it, you can [12:39] build local hackable agents with any [12:41] model or ship products on top of managed [12:43] cloud agents. The Cursor SDK handles the [12:46] harness, sandboxing, computer use, demo [12:48] videos, GitHub integration, etc. We're [12:50] going to talk about the Cursor SDK, what [12:51] it does, what people are starting to [12:52] build with it, but I want to make an [12:54] argument that it's actually part of a [12:55] broader phenomenon. In the past few [12:57] weeks alone, we've seen OpenAI make a [12:59] big update to their agents SDK. We've [13:02] seen Anthropic released Claude managed [13:03] agents. Microsoft released hosted agents [13:06] in Foundry, with CEO Satya Nadella [13:08] writing, "Every agent will need its own [13:10] computer, and with new hosted agents in [13:12] Foundry, every agent gets its own [13:14] dedicated enterprise-grade sandbox with [13:15] a durable state, built-in identity and [13:17] governance, and support for any harness [13:19] or framework." Cursor SDK is not the [13:21] same as all these products, but they are [13:22] all playing around in similar space. And [13:25] for the sake of simplicity, I'm going to [13:26] propose a new name for this category, [13:29] harness as a service, a new [13:31] infrastructure category where companies [13:32] sell access to their agent runtime, [13:34] which is the engine that turns an LLM [13:37] into something that can actually do [13:38] work, the same way AWS sells access to [13:40] compute and Stripe sells access to [13:42] payment rails. [13:43] So, before we can go to the next level [13:45] in explaining this, let's take a step [13:46] back and redefine harness itself. You [13:49] might have watched or listened to my [13:50] show about harness engineering a couple [13:51] weeks ago. It's a term that has been on [13:53] the rise all year. Akshay on Twitter [13:55] actually did a nice little summary. He [13:57] argued that the agent landscape has [13:58] evolved in three phases, writing that [14:00] the biggest shift in AI agents had [14:02] nothing to do with making models [14:03] smarter, but was instead about making [14:05] the environment around them smarter. In [14:07] phase one, the weights phase, he writes, [14:08] "Everything was about the model itself. [14:10] Bigger models, more data, better [14:11] training. Scaling laws told us that [14:13] progress equals more parameters. RLHF [14:15] and fine-tuning shaped behavior, and if [14:17] you wanted a better agent, you trained a [14:18] better model. In phase two, the context [14:20] phase, however, there was a new [14:21] realization. You don't always need to [14:23] change the model. You can change what [14:24] the model sees. That got us prompt [14:27] engineering, few-shot examples, chain of [14:28] thought, RAG. Suddenly, he writes, 'The [14:30] same frozen model could behave [14:31] completely differently based on what you [14:33] put in front of it.' Developers stopped [14:34] fine-tuning and started iterating on [14:36] prompts and retrieval pipelines instead. [14:38] But, he continues, 'Context windows are [14:39] finite. Long prompts get noisy, models [14:42] attend unevenly, and every new session [14:43] starts fresh with zero memory of what [14:45] happened before.' [14:46] Which gets us to the third phase, the [14:48] harness engineering phase. Akshay [14:50] writes, "This is where we are now, and [14:52] the shift is fundamental. The question [14:53] changed from what should we tell the [14:55] model to what environment should the [14:56] model operate in? The model is no longer [14:59] the sole location of intelligence. It [15:01] sits inside a harness that includes [15:02] persistent memory, reusable skills, [15:05] standardized protocols like MCP and A2A, [15:07] execution sandboxes, approval gates, and [15:09] observability layers. The model stays [15:11] the same. What changes is the task it's [15:13] being asked to solve. One concrete [15:15] example he gives is a coding agent asked [15:17] to implement a feature, run tests, and [15:18] open a PR. Without a harness, he writes, [15:20] 'The model must keep repo structure, [15:22] project conventions, workflow state, and [15:24] tool interactions all inside of a [15:25] fragile prompt. With a harness, [15:27] persistent memory supplies context, [15:29] skill files and code conventions, [15:30] protocolized interfaces enforce correct [15:32] schemas, and the runtime sequences steps [15:34] and handles failures. Same model, [15:36] completely different reliability.' [15:38] Importantly, he writes, 'Each phase [15:39] didn't replace the previous one, it [15:40] layered on top. Weights still matter, [15:42] context engineering still matters, but [15:44] the center of gravity has moved [15:46] outward.' [15:47] Now, I think intuitively, people have [15:48] experienced this to some extent. This [15:51] agentic explosion that we've experienced [15:53] in 2026 was not just the byproduct of [15:56] Opus 4.5 and GPT-5.2, although those [15:58] model updates were integral to it. It [16:01] was the combination of those models with [16:03] their harnesses, Claude Code, which was [16:05] the default harness for agent builders [16:07] throughout 2025, and more recently, of [16:09] course, OpenAI's Codex. The conversation [16:11] has matured to the point where in a [16:12] recent interview with Ben Thompson, Sam [16:14] Altman was asked, "How important is the [16:16] harness, the runtime around the model, [16:17] the tools, the state, to making agents [16:19] actually work?" Sam said, "Hard to [16:21] overstate how critical it is. I no [16:23] longer think of the harness and the [16:25] model as these entirely separable [16:26] things. Like, my experience of using [16:28] these, I am very aware of the fact that [16:30] I don't always know when I fire [16:31] something off in Codex and it does an [16:32] amazing thing for me, I don't know how [16:34] much credit" Ben Thompson fills in, "was [16:36] it the model that's amazing or the [16:37] harness that's amazing?" To which Altman [16:39] responds, "Yeah, exactly." [16:41] So, we now have these two very different [16:43] vectors of increasing AI capability. [16:45] There's the underlying models and the [16:47] changes there, but then there's also [16:48] improvements in the harness that [16:49] surround them. Of course, another part [16:51] of what has made this year feel so [16:53] different and what has empowered so many [16:54] different people is that we got an open [16:56] harness, even though that's weren't what [16:57] most people were calling it at the time, [16:59] that actually allowed us to build the [17:00] structure around agents to have them do [17:03] the things that we'd always imagined [17:04] agents being able to do. [17:06] I'm talking, of course, about Open Claw. [17:08] And yet, Open Claw was not [17:10] plug-and-play. You had to do everything [17:12] from picking the model to write the [17:14] system prompt to define the tool to wire [17:16] the agent loop, i.e., the part that [17:18] decides what to do next, that dispatches [17:20] tools, handles results, decides when to [17:21] stop. You had to manage context. You had [17:24] to handle errors. The user building [17:26] their Open Claw had to orchestrate [17:27] sub-agents when they needed parallel [17:29] work. They had to figure out how to [17:30] store state between runs. They had to [17:31] figure out where to deploy it, how to [17:32] monitor it, and if something broke, you [17:34] fixed it. If you wanted a new [17:35] capability, you built it. Every layer of [17:37] the stack was yours to assemble, [17:39] configure, and maintain. [17:41] I don't think it's a totally [17:42] inappropriate analogy to look at this [17:44] almost like the hobbyist era of [17:45] computing. [17:47] In a recent post on LinkedIn, Anders [17:49] Carlson wrote about the forgotten era in [17:51] computing. In the post, Anders talks [17:53] about how they started with a computer [17:54] in their school, a little mini computer [17:56] called an Alpha LSI, but quickly decided [17:58] that they wanted a computer for [17:59] theirself. They continue, "CompUKit [18:02] UK101. This was a kit in the true sense [18:04] of the word, not just a pre-produced [18:06] circuit board that you put in a case, [18:07] but an unpopulated circuit board, some [18:09] chips and other components, and a very [18:12] rudimentary assembly manual. You needed [18:14] a soldering iron, some basic tools, and [18:15] a ton of patience to build this thing. [18:17] This might be hard to grasp for people [18:19] using computers these days, but for a [18:21] few years in the 1970s, just after Ed [18:23] Roberts had released the Altair 8800 and [18:25] before the Apple II and that generation [18:27] of computers came around, there was a [18:29] short era of these truly hobbyist [18:31] computing. In other words, there was [18:32] this period, believe it or not, where [18:34] the way that a lot of people were [18:35] interacting with computers was having to [18:38] build and assemble them themselves. You [18:41] got a kit in the mail and you had to put [18:42] it together. Now, obviously, this period [18:45] didn't last long, and the number of [18:46] people who were willing to do this was [18:47] pretty small. It was the opportunity [18:49] that companies like Apple with their [18:51] Apple II seized upon in order to truly [18:53] get us into the PC era of today. [18:56] Now, at this point, you might be already [18:57] following the analogy that I'm about to [18:58] make, but effectively, the Open Claw era [19:01] of harnesses, you know, in the ancient [19:03] days, 2 and 1/2 months ago, was an era [19:05] where there were no pre-built harnesses [19:07] you could rent. You could either use the [19:09] basic harness plus model integration of [19:11] Codex and Claude Code, or you could [19:13] customize everything in Open Claw. What [19:15] we're getting now, with what I'm calling [19:17] harness as a service, is structurally [19:19] different. With these tools, the agent [19:21] loop is pre-built. The tool dispatch is [19:23] pre-built. Sandboxing is pre-built. [19:25] Streaming, error handling, context [19:27] compression, all pre-built and tuned by [19:29] teams whose full-time job is making [19:30] those layers excellent. You bring three [19:33] things: which model you want, what tools [19:35] the agent has access to, and what task [19:37] you're handing it. Everything underneath [19:39] is handled. This is obviously not a [19:40] shift in scale, but a shift in kind. And [19:43] just like the PC era didn't destroy or [19:45] undermine all those hobbyists who liked [19:47] having control over all the components, [19:49] it was an incredible democratizing force [19:50] that made it possible for people who [19:52] never would have built their own PC to [19:53] actually use a computer. The [19:55] productivity revolution of the 1990s [19:56] happened because users got Dell [19:58] desktops, not because more people learn [19:59] to assemble motherboards. That is the [20:01] promise of this new harness as a service [20:03] phase. [20:04] Now, the analogy is imperfect because [20:06] the actual democratizing force when it [20:08] comes to agents won't really for most [20:10] people be building their own agents. [20:12] It'll be using some set of prefabbed [20:14] agents as a service that they have a [20:15] very small number of knobs to customize. [20:17] And the applications that are powered by [20:19] agents that are more customer-facing. [20:21] What's weird and interesting and [20:23] different is that while on the surface [20:24] of it, tools like managed agents and the [20:26] cursor SDK seem like they're only for [20:29] developers. [20:30] Because we now live in the world where [20:32] agents handle the coding and can support [20:35] you around the infrastructure build, the [20:37] audience of people who can build with [20:38] something like the cursor SDK has [20:40] expanded dramatically. We've got this [20:42] entire new category of builders, of whom [20:44] I'm an example, who are not developers, [20:47] but for whom something like the cursor [20:48] SDK will matter. [20:50] Let's give an example of something that [20:51] someone built with this. [20:53] Jack Driscoll showed off a demo after [20:54] building with the SDK for a few days on [20:56] pre-release. The agent is effectively a [20:58] cursor agent embedded directly into [21:00] Gmail. From within his Gmail inbox, he [21:03] can share an email into chat, have that [21:05] agent read the email thread into [21:07] context, go off and do something, i.e. [21:09] edit some code, fix a problem, etc., and [21:12] stream the results right back into the [21:13] chat window. When someone asked why this [21:16] is different from just building with any [21:17] other SDK, Driscoll commented, "The [21:20] biggest difference in my opinion is that [21:21] cursor SDK isn't just calling LLM with [21:24] tools. It's exposing the same coding [21:26] agent runtime cursor already uses. Repo [21:29] context, edit, search, terminal [21:30] workflow, streaming status, model [21:32] choice, and local hosted execution. For [21:35] this demo, Gmail and chat is just the [21:37] intake and collaboration layer. Cursor [21:39] SDK is the part that can actually go [21:41] operate on a code base like a cursor [21:43] agent would. In short, the harness. In [21:45] other words, Jack is taking advantage of [21:47] this new harness as a service to build [21:49] exactly the agent wants in exactly the [21:50] environment he wants doing the exact [21:52] thing he wants." [21:53] Now, what's interesting is alongside [21:54] this announcement, we're seeing how much [21:57] harnesses can change the performance of [21:58] the underlying models. A new report from [22:00] Endor Labs found that GPT-5.5 operating [22:03] within cursor's harness set a new record [22:05] on their security correctness benchmark. [22:08] Benchmark test code for both [22:09] functionality and security. The cursor [22:11] plus GPT-5.5 combination achieved a [22:13] score of 23.5% on the security section, [22:16] which narrowly beat out the previous [22:17] leader, which was cursor and Opus 4.7 [22:19] with a score of 22.9%. [22:21] Both of these scores were a few [22:23] percentage points higher than Opus 4.7 [22:25] in its native Claude code harness and [22:26] GPT-5.5 in its native Codex harness. The [22:29] functionality test was arguably even [22:31] more stark. Opus jumped by almost four [22:33] percentage points from 87.2 to 91.1% by [22:36] switching to cursor, while GPT-5.5 went [22:39] from 61.5% to 87.2% by switching to [22:43] cursor. [22:44] Endor Labs wrote, "The key takeaway, [22:46] same model, same week, two harnesses, [22:48] two different functional results." [22:51] Alex Volkov from the Thursday AI podcast [22:53] confirmed that he found similar results [22:54] with an entirely different setup. [22:56] Testing the models on coding benchmark [22:58] WolfBench AI, he found that cursor's [22:59] harness had the strongest performance [23:01] for GPT-5.5 and was on par with Claude [23:03] code when running Opus 4.7. With the SDK [23:06] just launching yesterday, we're already [23:08] seeing a ton of people building MVPs for [23:09] agentic products with the new platform. [23:11] Tejas Vavery built a bug-catching agent [23:13] that can work on his production code [23:14] base and can see how the app is [23:16] performing on its own browser window. [23:18] This kind of agent could be a big step [23:20] to figuring out how to create fully [23:21] autonomous coding agents given that [23:23] human code verification is currently a [23:24] massive bottleneck. Vaver wrote, "It's [23:27] mainly about closing the feedback loop. [23:29] Right now, agents write code and hope it [23:31] works. They can run tests, but tests [23:32] don't catch everything, especially UI [23:34] behavior, integration issues, or flows [23:36] that depend on real browser state." He [23:37] explained that being able to actually [23:39] see the app is a huge change, allowing [23:41] for faster iteration, more confident [23:43] agent handoffs, and far less [23:44] verification work for humans. Robert [23:46] Brochery used the SDK to embed a cursor [23:48] agent in a Chrome plugin. The tool is [23:50] designed for IT triage, helping [23:51] non-technical users dump code from the [23:53] browser into a ticket instead of just [23:55] describing the bug and hoping for the [23:56] best. [clears throat] [23:57] Obviously, these use cases are just [23:58] scratching the surface, but they [24:00] demonstrate the power and potential of [24:02] freeing cursor agents from their IDE [24:04] container while retaining their runtime [24:06] environment. Basically, an entirely new [24:08] universe of agentic apps just opened up [24:10] now that developers don't need to start [24:11] from scratch. Now, what's very clear [24:13] when you watch the space closely is that [24:15] we are going to see massive amounts of [24:17] iteration and exploration around [24:19] harnesses. What's interesting to me [24:20] about harness as a service is the extent [24:22] to which those innovations don't just [24:24] end up producing better harnesses that [24:26] are embedded in the tools we use, but [24:28] actually allow people to build whatever [24:30] it is that they want to build leveraging [24:32] those harness innovations at a more [24:33] fundamental level. [24:35] If you want to check this out for [24:36] yourself, but you are not a developer, [24:38] what I would suggest is go to the cursor [24:40] announcement, click on the link to their [24:41] GitHub cookbook, drop it into either [24:43] Claude or ChatGPT, especially if it's [24:45] got context about your particular [24:46] projects, and ask it to give you some [24:48] ideas for how this new harness as a [24:50] service product could change the way [24:52] that you build or think about something. [24:54] Then, of course, when you realize that [24:55] it unlocks something that you've been [24:56] wanting to do forever, I give you full [24:58] permission to call in sick on Friday and [25:00] just dive all the way in. [25:02] For now, that's going to do it for [25:02] today's AI daily brief. Appreciate you [25:04] listening or watching as always, and [25:06] until next time, peace.