137: Agents Are the Opportunity—and So Are the Tools That Build Them | From OpenAI Dev Day｜Agent #6
Summary
OpenAI is compressing Agent building, deployment, evaluation and optimization into a standardized production line. Agent Builder orchestrates classifiers and if/else logic through drag-and-drop workflows; ChatKit handles frontend delivery; datasets, trace grading, prompt optimization and reinforcement fine-tuning form the feedback loop. Henry’s summary: OpenAI “trained itself in every martial art, and now wants to teach them to developers everywhere.” This lowers the barrier to enterprise adoption, but may also standardize away part of the performance moat that once belonged to Agent startups.
The core asset of Apps in ChatGPT is not a new UI, but operating-system-level distribution to 800M weekly active users. Unlike most GPTs in 2023, which were little more than “wrappers around prompts,” the new Apps SDK is built on MCP and adds OAuth-style authorization, external tools and interactive components, allowing apps such as Canva and Figma to deliver a fuller product experience. Revenue sharing remains experimental; the idea of a future 30% take rate is only an iOS-style analogy. Developers also face an asymmetry in data and retention: “It’s a bit like building a house on someone else’s foundation.”
Agent Builder exposes the real tension between OpenAI’s commercialization strategy and its AGI path. The AGI vision is for models to absorb hard-coded human workflows, while Claude Code-style general Agents improve directly as the underlying model gets better; Agent Builder instead turns workflows into diagrams. The trade-off is safety, explainability and something large enterprises can buy today. Henry called it “a very pragmatic pivot”: OpenAI wants autonomous Agents, but also needs current enterprise revenue to support a valuation of about $500B and even larger ambitions.
Agent Tooling could jump from a roughly $20B–$30B DevTools submarket to a long-term $200B–$500B market. The guest starts with a global software market of about $650B and DevTools at a low-to-mid single-digit share, then cites the view that software could expand to $10T as it absorbs services. If the tooling layer captures about 5% of the new market, it could grow 10–15x. The biggest opportunities are not tool catalogs, but products that control identity, communications, observability or evaluation and build data loops where “the more you use them, the higher the pass rate and the lower the cost.”
The tooling ecosystem is moving from “can call tools” to “can complete tasks,” with Composio as a representative case. It uses Agents to write and repair MCP servers, then positions Rube as “one MCP server to rule them all,” allowing Cursor to select tools from hundreds of servers based on the task. This eases context and token pressure while turning reliable execution into a product. MCP is better suited to high-success-rate critical workflows, especially those involving writes; browser use fills the long tail of sites without open APIs. The two are more likely to coexist than converge.
Voice has already produced a measurable infrastructure multiplier: LiveKit went from carrying about 1M voice calls per day to 20M within a year. It does not primarily make voice models; it provides real-time audio/video, turn detection and orchestration. Its customers include Character.AI and Grok, and it says it supports roughly 25% of 911 traffic and helps save one life on average each week. Speech-to-speech is seen as the end state, but the STT→LLM→TTS cascade still has a clear commercial runway because it offers more control over guardrails, cost and behavior.
Evals and memory are not ancillary features; they are the control layer that moves Agents from demos into production. Even a large customer may release a voice Agent after engineers test only 3 or 4 calls because labeled data is expensive, teams cannot agree on evaluation sets, and subjective, complex tasks are harder to verify than coding or math. OpenAI’s $1.1B acquisition of Statsig shows that A/B testing, phased releases and metric loops are being built in. Letta’s “sleep-time compute” represents another path: when no one is interacting with the Agent, it spends tokens organizing long-term memory and continuously converting state into reusable capability.
Deep dive
1. AgentKit Turns the Agent Development Loop into a Set of Standard Components
Henry’s summary of the launch was that OpenAI had “trained itself in every martial art, and now wants to teach them to developers everywhere”: AgentKit covers build, deploy, monitor, evaluate and optimize, rather than offering only a model API.
Agent Builder is a visual orchestrator. On stage, the team built a production-grade Agent from scratch in 8 minutes by dragging and dropping a classifier, then adding if/else branches that called different Agents. At its core, this remains a workflow “crafted entirely by humans.”
ChatKit fills in the frontend UI SDK, allowing developers to embed chat components. After launch, datasets, trace grading, automatic prompt optimization and reinforcement fine-tuning take over, forming a loop from user interaction to model improvement.
A trace is the complete execution path from a user request through success or failure. Trace grading can use rule functions or an LLM-as-a-Judge to label and score individual steps and locate problems.
2. The Value of a Visual Builder Is Enterprise Adoption, Not Technical Novelty
Manqi asked whether it was similar to Dify, Coze, n8n or Zapier. Henry’s answer was broadly yes: OpenAI’s UI/UX may be more polished, but it has not yet created a meaningful mechanism-level difference. That is also why some developers were disappointed.
Silicon Valley developers are already used to building Agents directly with Claude Code and Codex. Drag-and-drop does not materially change their workflow; the real incremental users are less-technical business staff inside enterprises who need processes that are safe, legible and easy to explain.
That also means some Agent companies will see their differentiation standardized. Continuous evaluation and tuning may once have been the moat that let one company build a better Agent; OpenAI is now packaging datasets, trace grading and reinforcement fine-tuning together.
3. Agent Builder Has a Directional Tension with the Autonomous AGI Path
Henry preserved the critics’ core objection: the AGI path aims to “eat” the steps humans have hard-coded, allowing a stateful Agent to plan autonomously and use tools continuously. Agent Builder instead draws the workflow as a graph and hands it to the model, making the two approaches “even somewhat opposite.”
Claude Code represents the general-Agent path: let the Agent run by itself as much as possible and draw fewer workflow diagrams. When the underlying model improves, the entire Agent benefits directly. Agent Builder instead admits, “I may not be able to wait for future models to get better,” and meets enterprise needs today.
The product is “not sexy” to researchers pursuing AGI, but it is safe, understandable and deployable for large customers. That is why Henry believes “it should ultimately get people to pay.” It is not the technical end state, but it may be the faster route to commercialization.
Manqi noted that OpenAI is pursuing both paths, with a code-based Agents SDK as well. Henry’s response was that Sam Altman may “want everything”: as the organization grows, it can tolerate forks, while valuation and revenue pressure demand pragmatic products.
4. Apps in ChatGPT Are Finally More Than a Prompt Store
The central problem with custom GPTs in 2023 was that many products ended up as little more than “a prompt template with a custom logo.” They could not reliably call third-party services, preserve state, send notifications or execute complete workflows, leaving the GPT Store “long on thunder and short on rain.”
This time, the Apps SDK is built on MCP and adds OAuth-style authorization, external tool calls and interactive UI. Combined with evals, versioning and ChatKit building blocks, developers finally have the toolkit to build a complete app experience rather than “a wrapper around a prompt.”
Canva makes the before-and-after clearest. The old version could also generate images; the new one can connect and authorize a Canva account, bringing the user’s existing assets into the generation workflow. Launch partners including Canva and Figma also raise the quality bar for the first wave of apps.
5. 800M Weekly Actives Create Operating-System-Level Distribution
Henry sees the most immediate benefit as distribution: ChatGPT has about 800M weekly active users, so a high-quality app can reach a large audience immediately without solving cold start on its own. For startups that need security and compliance credentials, the platform may also lower the trust barrier to entering large enterprises.
Compared with the lack of monetization examples for GPTs 2 years ago, the developer incentive may be more direct this time. But pricing, revenue sharing and advertising remain unclear at launch. Henry mentioned a possible 30% take rate in the future, while stressing that this was only speculation based on the iOS model.
“ChatGPT could become an operating system” is not just a slogan. The Apps SDK puts interaction components resembling native apps directly inside text conversations, allowing users to authorize, operate and complete tasks without leaving ChatGPT.
6. OpenAI Is Refining AgentKit Through Its Own Enterprise Practice
Henry believes OpenAI will not abandon B2B. Anthropic and Google are aggressively competing for enterprise and developer business, while OpenAI has also formed a Future of Work team focused on customer service, contract review, data analysis and sales-lead conversion.
The “OpenAI on OpenAI” approach is to run the full sales, closing, customer service and customer success workflow internally with AI first, then distill the experience into tools. AgentKit is therefore not just a lab product; it is an externalization of internal production practice.
One concrete figure: OpenAI receives about 13,000 sales leads per month, but its human staff previously provided relatively personalized replies to only about 1,000. AI has sharply expanded coverage efficiency.
Customer-service Agents query internal knowledge bases and policy repositories; trace grading during service may feed back into those repositories. Henry’s causal chain is simple: “The more Agents you build yourself,” the better the evaluation and development tools naturally become.
7. Enterprise Data Is Fuel the Model Race Cannot Forgo
Henry’s B2B argument is not only about revenue. Model intelligence is multidimensional, and without enterprise-scenario data it is difficult to improve across different business dimensions. Those gaps ultimately hurt ChatGPT’s productivity use cases as well.
OpenAI is investing simultaneously in foundation models, an app for Sora 2, the developer platform and enterprise Agents. That has prompted debate over whether these commitments could slow the pursuit of AGI, but the company has already grown to several thousand people and is choosing parallel expansion rather than a single-track bet.
On the “too big to fail” question, Henry agreed that OpenAI may already be too large to fail. But a valuation of about $500B may not satisfy Sam Altman; if the company wants to move up another order of magnitude, it will still need real commercial value to support it.
8. Platform Benefits Come with Three Risks: Data, Retention and Self-Competition
The first asymmetry is data. App context and user behavior remain largely inside ChatGPT, while developers may receive only limited invocation data, making it difficult to build direct user relationships, analyze behavior or optimize retention. “It’s a bit like building a house on someone else’s foundation.”
The second risk is the platform entering the market itself. OpenAI already covers B2B workflows including sales, conversion and customer service, and it controls the full conversation history. Opportunities that emerge on the platform could be built internally or absorbed directly.
Manqi used Google as a comparison: developers also worry that Google will copy successful apps, but generally assume a large company’s product velocity will lag a startup’s. OpenAI executes faster, making it more intimidating; that could in turn weaken its neutrality as a platform.
Agent Builder currently supports essentially only OpenAI models, while evals can bring in third-party models. Henry believes that if Apps in ChatGPT is to become a major platform, it should eventually open up model and tool choice and narrow its own role to the platform layer.
9. Traffic May Concentrate Around ChatGPT and Gemini
Henry believes the eventual number of platforms depends on how many “large traffic entry points” exist. The most prominent today are ChatGPT and Gemini. Gemini is reportedly ahead of ChatGPT in total users if the entire platform is counted, though that figure may include usage through Gmail and other surfaces.
Gemini’s catch-up momentum still deserves attention, and AI Studio’s development features continue to improve. OpenAI, meanwhile, used Dev Day to fill in the Agent development chain. Both sides are moving quickly.
Anthropic occupies a different position. Claude is not yet an equivalent mass-market entry point, while MCP is an open standard that every model can adopt. Henry therefore does not believe MCP automatically gives Anthropic a strong competitive advantage, even though it materially strengthens the ecosystem as a whole.
10. Agent Tooling Is the Body and Coach of a Virtual Digital Person
Henry defines AGI as a kind of “virtual digital person.” In the past, humans learned to use mice and keyboards to adapt to machines; in the future, computers will adapt to people, and human-computer interaction will increasingly be based on natural conversation—“everyone-to-everyone interaction.”
Large models are only the intelligent brain. LangChain is the skeleton for coordination and scheduling; MCP is the left hand for tool use; browser use is the right hand for operating websites; ElevenLabs is the mouth; and LiveKit provides the real-time perception infrastructure needed to hear and see.
Eval companies such as Braintrust and Galileo are more like coaches. They do not directly form the body, but continuously observe Agent performance, identify failures and drive improvement, moving Agents from something that can be demonstrated to something that can execute reliably.
11. Six Model Waves Gradually Created the Full Tooling Ecosystem
The first wave began with ChatGPT and GPT-3.5 at the end of 2022. Developers realized that conversation alone was not enough: they needed external data, context management and multiple chained LLM calls. LangChain-style scaffolding emerged, and these products were then still called “large-model applications.”
The second wave was OpenAI Function Calling in June 2023. Models began calling peripheral tools based on context. Henry even said, with some exaggeration, that early Agent development spent “80% or 90% of the time” writing a good tool.
In November 2024, Anthropic released MCP, allowing tools to be reused across different models. Existing tool providers began converting into MCP server providers. Henry called it “a gift and an act of charity” from Anthropic to the community and the AI ecosystem.
The third and fourth waves came from GPT-4o Advanced Voice Mode in May 2024 and Claude Sonnet 3.5 in June 2024. The former pulled LiveKit forward; the latter launched Cursor-style AI coding and helped spawn secure code sandboxes such as E2B and Daytona.
12. Reasoning and Computer Use Pushed Agents from Scripts toward Autonomous Execution
The fifth wave was o1-preview in September 2024. Reasoning allowed Agents to plan multi-step tasks and tool sequences on their own, while making reinforcement-learning fine-tuning services and complex-task evaluation companies such as Braintrust and Galileo more important.
The sixth wave was Anthropic computer use in October 2024, followed by OpenAI Operator. Browser operation created cloud-browser infrastructure companies such as Browserbase and Anchor Browser, as well as higher-level APIs that accept natural-language tasks such as “order me some takeout.”
Henry admitted that forecasting the next 3 to 6 months is “very, very difficult.” What seems relatively certain is that every model company will strengthen reasoning and tool use, voice will continue to attract investment, and multimodality will become more integrated. Nano Banana-style image-editing models came up as an example.
13. Capital Is Betting Simultaneously on the Mouth, Skeleton and Coach Layers
Naomi’s first signal was an employee secondary sale at ElevenLabs of about $100M at a valuation of roughly $6.6B. Only 9 months after the previous financing, the valuation had approximately doubled, showing that adoption and monetization of voice infrastructure are both accelerating.
LangChain is closing a financing round of about $100M led by IVP at a valuation of roughly $1.1B. Capital continues to back Agent frameworks and developer infrastructure rather than chasing only end-user applications.
OpenAI acquired Statsig for about $1.1B, turning A/B testing, phased releases and data-metric loops from external tools into built-in components. Both observers noted that more and more people who used to work on A/B testing are moving into evals.
14. Composio Moves MCP from a Tool Directory to Task Completion
Naomi’s core view is that the industry is moving from “the era of tool calling” into “the era of task completion.” Composio does not merely provide a large number of MCP servers; it also tries to ensure that Agents execute tasks more reliably, functioning like a skills layer or operating system.
On one side is an MCP platform for developers to integrate programmatically; on the other is the prosumer product Rube. Its slogan borrows from The Lord of the Rings: “One MCP Server to Rule Them All.” A Meta MCP selects the correct server based on the task.
Henry recalled that Cursor had only about 3 MCP servers lit up at the same time, while explicitly noting that the specific number might need confirmation. Each server exposes multiple tools and many parameters, and putting everything into context consumes a great deal of tokens.
Tool count involves a trade-off. Complex products such as Gmail may need to expose many capabilities, but when two functions are too similar, the model becomes confused about “which one to call,” reducing execution success.
15. Composio’s Moat Comes from Building Agents First, Then Having Agents Build Tools
Composio was founded in India around July 2023. It initially wanted to build an integration-heavy Devin-like product that used AI to generate third-party API integration code. Model accuracy was not yet good enough, and the team found a more fundamental problem: Agents could not call tools reliably.
Before MCP was released, the company pivoted away from integration automation and focused on packaging general-purpose tools into a skills layer that LLMs could call directly. Today, the production and improvement of high-quality servers is itself automated by Agents.
With limited resources, the team built its reputation through Twitter demos, feedback from Reddit and Discord, and hands-on debugging of integration requests. It later raised a Series A led by Lightspeed and accumulated customers including ClickUp and Glean.
The team now has about 30 people and moved from India to San Francisco to stay close to Agent customers. Naomi said she had not yet met a non-Indian employee; the team is looking for the “scrappiest” and “hungriest” people rather than deliberately localizing hiring after moving to the US.
16. Anthropic Has No Need to Maintain a Full Composio Itself
Manqi’s question was: if MCP came from Anthropic, why not build the integration marketplace itself? Naomi’s view is that maintaining fragmented third-party APIs would not strengthen Anthropic’s core moats in model capability or safety, and would instead consume substantial manpower.
Naomi added that the underlying platform also wants to remain relatively neutral. An open protocol that expands the tool ecosystem available to Claude may be enough; Anthropic does not necessarily need to own the engineering maintenance and customer support revenue for every server.
Composio exists in the gap created by that division of labor: model companies define the interface, while specialists handle fragmented third-party API maintenance, error handling and reliable execution in tool calls.
17. MCP and Browser Use Will Coexist Long Term Because Their Reliability Differs
On benchmarks such as OSWorld and WebArena, the guests said the best desktop or browser models have success rates of about 60%–70%. Henry posed a medical-onboarding question: if a critical workflow succeeds only 70% of the time, “I think that is completely unworkable.”
In the near term, high-risk workflows and write operations will therefore depend more heavily on reliable tool use. As reasoning and tool use improve, Agents may stop “letting a child swing a broadsword” and move from today’s common read-only research into genuine execution.
Browser use matters for the long tail not covered by MCP. Many boutique hotels in Japan are not connected to OTAs; one team uses a browser to iterate through date combinations and scrape room availability for search engines. These nonstandard sites are not worth building formal APIs for one by one.
Henry’s end-state metaphor was: “The ultimate goal of ChatGPT is WeChatGPT.” If it becomes a super app that handles payments and all kinds of services, service providers may no longer be deciding whether to connect to an LLM interface; they may be facing “a tide too powerful to resist.”
18. The Voice Market Is Already Showing Infrastructure-Level Growth
Henry said there are “one trillion” phone calls worldwide every day. His point was that many are business calls, customer service interactions and reservations that AI can handle, alongside 2 fast-growing categories: personal assistants and emotional companionship.
A year ago, LiveKit carried about 1M voice calls per day; by the time of recording, that had reached about 20M, a 20x increase in a year. It began during the pandemic as WebRTC infrastructure for real-time audio and video, with little initial connection to AI.
After GPT-4o adopted its transport infrastructure, LiveKit caught the first AI wave and expanded upward from the base layer. Its SDK can now help build voice Agents and handle turn detection and orchestration.
This is different from ElevenLabs. The latter initially looked more like a voice-model company, while LiveKit is real-time communications infrastructure for multiple models and scenarios—similar to Agora, but with an Agent platform as well.
19. The 911 Case Shows Real-Time Communications Go Beyond Cutting Customer-Service Costs
LiveKit says it carries about 25% of 911 traffic. After receiving a text message, callers can open a link and transmit their phone’s video and audio to the dispatcher in real time, allowing the dispatcher to see the scene rather than rely only on a verbal description.
In emergencies, the system can also bring in a remote CPR coach who guides the caller based on the live video. Henry said the infrastructure can “bring one person back from the brink of death every week on average.”
The voice experiences for Character.AI and Grok also run on LiveKit, but “powered by LiveKit” refers to communications and Agent infrastructure, not the source of the voice model.
A partnership that was about to be announced at recording was Salesforce: its Agentforce customer-service voice Agents will run on the LiveKit Agent platform. The platform also supports models from OpenAI, ElevenLabs, Cartesia and others.
20. Speech-to-Speech Is the End State, but the Cascade Is Still the Best Practical Answer
A cascade voice Agent performs STT first, sends the text to an LLM and outputs the result through TTS. Speech-to-speech, as in OpenAI GPT Realtime, returns voice directly from voice input. Both architectures can run on LiveKit.
Henry’s central judgment was clear: “The end state will definitely be speech-to-speech.” End-to-end models can keep laughter, emotion, intonation and mid-conversation language switching inside the context, making the interaction more human.
But the cascade is more controllable today. Developers can add guardrails and decision logic at the string stage. Hospital information collection, for example, does not require AI to pretend to be human; a degree of mechanical quality may actually be more appropriate.
Investment in voice is still increasing. David AI, which provides voice data to top labs, recently completed a Series A of about $25M. OpenAI’s GPT Realtime API became generally available last October; this launch is version 2.0.
21. Voice Competition Is Consolidating among Several Players, with Anthropic Not Yet at the Table
The guests observed that OpenAI, Google and Grok are all increasing their voice investments, while Anthropic “may not be at the table,” appearing more focused on going all in on coding.
The new GPT Realtime improvements center on more human-like voices and expression, better adherence to prompts such as “warm” or “professional,” and the ability to understand laughter, emotion and language switching mid-sentence.
Developers in China widely use MiniMax and Doubao voice models, while podcast-generation Agents are also proliferating. Henry believes Chinese voice-generation models are highly competitive, noting that MiniMax ranks near the top of Artificial Analysis and may even be ahead of ElevenLabs, though its overseas distribution remains limited.
22. Agent Memory Includes at Least Four Distinct Product Needs
Episodic memory lets a customer-service Agent remember that it “already tried solution X last week,” avoiding repeated actions and preserving continuity. This is not simply storing chat history; it is extracting experiences useful to the next task.
Procedural memory stores the path used to complete a task. A DevOps Agent might remember that the previous deployment failed at step 3, allowing it to avoid the bad step and reuse an effective path. This is useful for enterprise automation and multi-step Agents.
Semantic memory stores facts and rules. A shopping Agent might query a discount-rules database to reduce hallucinations about promotional conditions.
Role or personality memory preserves the conversational style of products such as AI boyfriends and AI girlfriends, preventing them from going out of character. Enterprise Agents may not need personalities, but they may rely on procedural and semantic memory at the same time.
23. Letta Turns Idle Time into Learning Time with Sleep-Time Compute
Letta was founded by 2 Berkeley PhDs to help developers build stateful Agents. Its signature concept is not test-time compute, but “sleep-time compute.”
Henry’s analogy is that people hold many meetings during the day, then reflect and process them at night to form insights. Letta allows the system to “spend tokens while sleeping” when no one is querying it, organizing past interactions and distilling the most useful knowledge.
Developers define a memory schema, such as identity, preferences and birthday. Letta reads conversations, updates the fields automatically and inserts relevant memories back into context when needed in the future, so it necessarily needs access to interaction content.
In the early days, vertical Agents will build memory systems independently, with “every group showing its own ingenuity.” Henry expects standardized products to emerge once the methods mature and take over the dirty work, although data sensitivity and how much content customers are willing to expose will remain scenario-dependent.
24. Memory Ultimately Remains Part of Context Engineering
OpenAI’s own memory feature is also evolving. A sentence such as “My mother made my favorite tiramisu for my birthday” may now be split into a name, a birthday, a preference for tiramisu and a specific event rather than stored as one complete sentence, making later retrieval more accurate.
Model capability and long-term memory remain 2 separate paths. Models can expand their context windows, but context is “not a free lunch”: cost and performance both deteriorate as length increases.
Henry summarizes context engineering as deciding “what information should enter the context and what should not.” The inner loop handles information needed for the current generation; the outer loop continuously improves the ability to select the right information over the long term.
Memory must enter context when it is used, so it belongs to context engineering. For current models, long-term memory is more like an “external hard drive” than a cognitive function already built into the model’s parameters.
25. The Evals Paradox: Everyone Says They Matter, Almost No One Does Them Properly
Henry’s exact point was: “Most people agree that evaluation, or evals, are important, but most companies today do not actually do evals well.” There is a large gap between production practice and verbal commitment.
A large LiveKit customer had a customer-service Agent that tended to hang up too early. The team added a single sentence to the prompt—“Don’t hang up so early”—then had engineers make 3 or 4 test calls. Once it felt as though the Agent was hanging up slightly later, they deployed directly to production.
A more reliable process would run the new version through hundreds of calls before launch, covering different scenarios and datasets, then analyze logs and confirm that the metrics had genuinely improved. The case also shows why eval components matter inside a voice-Agent platform.
26. The Difficulty of Evaluation Lies in Both Data and Organizational Consensus
Evals first require human labels or carefully constructed datasets. The more complex the task, the higher the cost. More difficult still, development teams may not agree that a dataset represents real-world quality.
Henry has seen teams spend a long time building an evaluation set, only for the results to show that a new version improved while team members said, “The dataset says I got better, but I feel like it got worse.” A metric without consensus cannot become a release gate.
Coding and math are well suited to reinforcement learning because answers and test signals are relatively clear. Subjective, complex tasks may take “more time to evaluate than to perform,” which is the central difficulty of general-Agent evals.
OpenAI is linking datasets, trace grading, prompt optimization and reinforcement fine-tuning, then acquiring Statsig. The direction is to turn evaluation from a one-off test into a continuous product and training loop.
27. The Long-Term Agent Tooling Market Could Be an Order of Magnitude Larger than Traditional DevTools
The guests’ baseline was a global software market of about $650B per year, with traditional DevTools accounting for a low-to-mid single-digit share, or roughly $20B–$30B.
The industry projection cited in the discussion is that AI could convert labor services into software services and push the software-market ceiling toward $10T. The key change is not dividing up existing software budgets, but reaching the much larger services economy.
If Agent Tooling captures about 5% of the new market, its long-term scale could reach $200B–$500B, roughly 10–15x the existing developer-tools market. This is a ballpark estimate, not a short-term revenue forecast.
28. $10B Companies Will Emerge from Critical Workflows and Data Loops
Traditional benchmarks include Okta at about $2B in annual revenue, Twilio at about $4B and Datadog at more than $2B. The Agent economy will rewrite the objects and workloads of identity, communications and observability.
New demand includes Agent identity and scheduling, behavioral trace replay, real-time low-latency communications, and code auditing, regression testing and secure execution driven by autonomous coding. Braintrust, LiveKit and Composio each control a different node in this stack.
Not every incumbent will be disrupted. The real paradigm opportunities lie in new behavioral layers that traditional metrics such as CPU and memory cannot describe—for example, explaining why an Agent chose a particular tool and where it went off course.
The key standard from the guests was whether a company can build its own data loop and control a workflow node: can the product become “better with use, higher-pass-rate and lower-cost”?
29. San Francisco’s Ecosystem Uses High-Frequency Feedback to Push Tool Companies toward Product-Market Fit
AGI House brings top-lab researchers, engineers and Agent Tooling companies together through research-paper reading groups, technical discussions and large hackathons. Startups get marketing as well as first-hand product feedback.
The investment case for Composio came from multiple Agent Build Days, where community feedback suggested that it was relatively reliable. Community signals arrived before formal financial data, allowing investors to observe bottom-up adoption.
San Francisco is becoming active again, with OpenAI and Anthropic offices both in the city. The guests said local rents appeared to have risen about 20%–30% at the time. Many teams formerly based on the Peninsula or overseas are also moving closer to customers and talent.
30. Chinese Founders’ Product Sense and Execution Are Forming a Complementary Advantage in Silicon Valley
Naomi said there were “a lot” of teams moving from China to the US that year. About 20%–30% of founding teams across the previous 3 or 4 HF0 cohorts had Chinese backgrounds. Each cohort has about 10 companies and accepts only serial entrepreneurs; the list is not public, but investors are selectively invited to Demo Day.
She observed that these teams often start companies in China, then use US accelerators to build connections and growth channels. Compared with YC, their teams, products and valuations may be more mature.
Henry believes Chinese founders have an advantage in B2C product sense and execution. A vertical Agent usually has only 2 moats: domain expertise and “pure execution.” The latter is a traditional strength of Chinese teams.
Silicon Valley offers a strong B2B operating environment and clear acquisition exits, which concentrates founders there. Chinese teams are better at B2C, including hardware, and may bring ambitions for “large-scale consumer applications” that are scarce among local investors, creating a US-China complementarity.
31. RL’s Second Spring Is Creating 2 Types of New Tool Companies
Henry calls one of the technology community’s other major threads “RL’s second spring.” Its starting point can be traced to about 9 months earlier or to the release of DeepSeek; reinforcement learning is no longer confined to foundation-model labs.
The first type of company builds RL environments and sells them to labs training foundation models. The second provides RL-as-a-Service, helping large enterprises improve internal models and Agents in specific business scenarios.
That closes the loop on the entire discussion: the stronger model reasoning becomes, the more complex the tasks Agents can handle; the more complex the tasks, the greater the need for tools, environments, memory, evaluation and reliable execution. The tools that build Agents are therefore an opportunity alongside the Agents themselves.