Anthropic's Platform Grew Three Layers in a Year — and Tokens Now Have Jobs

When Katelyn Lesse joined Anthropic about a year ago, the developer platform was basically the Messages API: a stateless call, plus SDKs, docs, and a console. MCP had already shown up as a standard. What had not shown up yet was the rest of the stack now running on that platform — managed agents, self-hosted sandboxes, skills, memory, and a layer they are starting to call strategies, where a token is no longer homogeneous fuel. Each token can be given a job: execute, advise, dream, or get graded and try again. Sonya Huang, hosting Sequoia Capital's Training Data with Lauren Reeder, opens by calling it one of the most important developer platforms in the world, if not the most important. Lesse and Angela Jiang lead the team that builds it. Externally it is the APIs and tools every developer uses to reach Claude. Internally it is the layer Anthropic's own apps sit on.

The team has two north stars, which Jiang admits sounds wrong — there should only be one — until you notice they sit in separate solar systems. Internally, the job is to give Anthropic's product teams as much leverage as possible to ship "AGI-pilled" products, fast, on a reliable platform. Speed is intentional. Externally, the true north is to give any builder the tools to work with Claude on whatever they want to build. That boils down to putting the platform wherever the business is, which is why they spend a lot of time integrating with hyperscalers such as AWS and Google. Sometimes that comes as primitives and higher-order abstractions. Sometimes it comes as standards. Skills and MCP, she says, are things Claude needs in order to be useful, so they can just give them out to the rest of the ecosystem and work with everyone.

Under both stars sits a rule they try not to break: internal teams and external developers get the same primitives. Plenty of platform businesses bifurcate the two. Anthropic tries to keep them equal. The thesis is that model capabilities are growing so fast that no form factor lasts. Two years ago everyone said everything was chat. Now everyone says forget chat, it is agents. There will be another form factor after that. They do not think they are the only people who can find those shapes; more democratization, they argue, is how the market actually surfaces them. Lesse's practical version is dogfooding inside while opening early access outside, so they do not over-index on internal requirements. Internal users and external users both have specific needs. Fall into one camp and you miss the other.

As models got better at running longer and holding more context, Lesse says, customers kept solving the same problems Anthropic was solving internally: agents that succeed in a long-running, even remote, context without a human in the loop. Infrastructure is hard. How do you spawn sandboxes with the right governance and security, spin them up and down, store transcripts so a session can pause and resume? Harness engineering is the other pile: prompt caching, context-window management, getting more intelligence out of the model, watching cost. The ethos of the platform now is to package those primitives so that if you want to offload some of that, you can, and stay focused on the problem you actually came to solve.

Jiang's oversimplification is a three-layer cake. At the bottom is knowledge: knowledge about the model, about what it needs, and about how to actually do something with Claude. The primitives there are more baked — specific shapes and parameters on the Messages API that try to express how Claude thinks, how it respects parameters, how it does tool calls — then standardized tools, then bits of context you can drop in at different moments, concretely skills and memory. MCP lives here too. The next layer is execution. Knowing is not enough; you need Claude to do work, edit files across systems, produce output. That requires infrastructure. The abstraction is a low-level harness plus managed infrastructure. Today's high-level product for that is Claude Managed Agents.

On top of that she sees a coordination layer, still more inkling than product. They have started to think of things called strategies, almost a meta-harness. The true low-level harness is designed for execution. The next one starts from the idea that tokens are not really fungible: this token is advising, this one is executing, this one is dreaming. You compose orchestrated strategies that sit on top, because you still have to execute, and execution still has to know what to do. Everything should ladder. Look at the roadmap and project forward, she says, and you will see them move more and more from knowledge to execution to coordination in the abstractions they put out.

Who reaches for which layer depends on the user. AI-native startups tinkering at a low layer grab the primitives. Everyone else — classic enterprises, or startups whose philosophy is not hill-climbing the agent stack but stringing together workflows and providing unique user value — reaches for the packaged offerings. It is not their core competency. Managed Agents, Lesse later notes, is called higher-order but is not that high: you still define custom tools, skills, system prompts, MCP servers. The knobs are still there. The end state they want is telling an agent the outcome you want and the budget you will spend, then ready, set, go.

Huang asks the title question: open ecosystem or walled garden, and which products they need to own first-party. Lesse answers on the cake. On execution, inside Claude Managed Agents, they are not precious about whether you run the work on their infrastructure. They launched self-hosted sandboxes and partnered with Modal, Vercel, Cloudflare, and others — even Amazon's new micro-VMs — so you can plug those in as a first-class offering. They launched MCP tunnels so you can call MCP servers behind a firewall. Whether it runs on Anthropic's infrastructure or somebody else's is actually not important to them. What is important is the architecture of how you put agents together so they are powerful, reliable, and scalable. They have strong opinions on that. Conform to the interfaces and plug things in.

Jiang's step-back is electricity. Before electricity you had a candle, and you could only do so many things. Electricity became a transformative utility because you could wire it into everything, everyone could access it, and there were standards for how to plug in. That is not something anybody can do by themselves. You always have to work with the ecosystem and with partners. They want a company to be able to get created, build whatever products it wants, build agents if it needs to, and have those agents plug into other agents — some Claude, some not. Transactability across the board. Standard-setting is part of that: skills and MCP at the builder layer, and, at a higher layer, how the industry treats safety together. Nobody wants negative activity on their service. Cyber is the example. Protect your own systems; partner on critical infrastructure and fraud.

The ecosystem is already composing in ways they did not ship as a product. Lesse describes a customer who built some agents on Managed Agents and others on other models and platforms, optimized each, then had a galaxy-brain moment: expose an MCP server on top of one agent so another agent can call it as a tool. Anthropic sat down with them, worked through it, and it worked. Agent calling agent is already happening. Another cluster of customers is fighting old software. Healthcare companies tell them the systems do not even have APIs; that is a dream. So they use computer use to automate and connect, sometimes a laptop running a pile of jobs to auto-generate interfaces their agents can consume. Jiang's read of the field: the most exciting innovation right now is at the context and connectivity layer. Customers have built custom software-development lifecycles and automated entire back offices; the nuance of how they stream context in is where the inventiveness shows. It does not express itself as a new product form. It expresses itself as being maximally useful.

When Anthropic builds first-party products, Jiang says, it uses two frames. The first is form-factor hunting. Form factors are not static. Something awesome for one year of AI development will probably not be awesome the next. Internally they ask whether a thing is AGI-pilled enough; they also build something, watch it be cool for a year, throw it away, and try again. They tell platform users the same. Claude Design is an example people might squint at as a move into design as a vertical. Look closer and two decisions show. First, offload more and more to Claude: talk to it, let it figure it out, discourage a little of the manual editing. Second, ask whether code can replace the classic WYSIWYG path — generate the design as code rather than integrate against a design system. Early experiments suggested it could. A lot of other internal projects live in the same bucket. The labs team tries them; they are super cool for two weeks; they move on; they often never ship.

The second frame is TAM. They are a business. They look for areas with reasonable agentic operations, and they orient toward work that is token hungry: after one turn, do you feel done, or so glad you did that thing that you want to do more of it? Coding is the obvious case. You finish a turn, you feel unlocked, you build more. Other services complete a job and you move on; those interest them less. They also pick business functions where the buyer wants workflow help. They have been transparent about verticals: finance and legal, with the right context and tools in a useful form factor. For finance you can build on the Messages API, or on Managed Agents, or as a connector inside a product. They recently launched Claude for Financial Services — packages of skills you can use in their product or in other people's — and cookbooks on how to use Managed Agents for those tasks. They put the pieces out and see where people run.

Sometimes they package the pieces themselves. Claude Tag is the example. They had watched the industry do this: Shopify with River, Block with BuilderBot, companies building an internal agentic platform, stuffing it with the right context, and making it reachable from Slack. Tag is that packaging — how Anthropic does it internally, ready to plug in. Huang asks what people misunderstood, given the ruckus that it was just a Slack bot. Jiang's answer is that past products trained people to attach to the UI. With Tag, you tag it in Slack; that is the interface; that is not the important part. The important part is the context engineering and architecture underneath so that it just works. It should feel like a coworker who comes into your channel, chats, is proactive, figures out what is useful, and gets stuff done. For non-technical people that is a huge unlock. How do I submit an expense report? Used to mean manager, spin-up buddy, running all over. Now you create a channel and @Claude.

Andrej Karpathy, she notes, put it well: it is an org-level harness. You could construct it from the APIs, but you would do a lot of experimentation yourself. The futuristic bit is that a lot of the complexity is an iceberg. The stuff underneath is becoming the harder and more useful part. The bit above the water — the interface — can constantly swap. Today a lot of businesses collaborate in Slack. Others are in Teams, a WhatsApp group, texts, still email. Agents can take up the same form factors humans already occupy. It is almost a boring take, she says, and she feels it is the most forward one: you want the agent to be another person helping you, intelligent, full of context, always on.

Huang asks for harness and context best practices. Lesse's list is deliberately boring. Prompt caching: do it, you will save a lot of money and tokens. Keep the context window clean; clear old stuff out. Sometimes call tools programmatically so you do not pull everything into the window. Then evals — she is surprised they got this far before anyone said the word. Those details live in the lower-level harness, the nitty-gritty they already packaged into Managed Agents. The interesting innovation, she says, is moving up to strategies. You can spend a given token on executing. You can spend that same token reflecting on past agentic sessions and writing learnings to memory so the next agent does a better job. You can advise with a bigger model so a smaller one can execute. You can execute, execute, then let a grader in: did you do a good job? No. Try again. There is not necessarily so much juice left to squeeze at the caching-and-clearing layer compared with a layer higher than that.

Jiang's explanation is generational. Two years ago a lot of the harness was a scaffold to take the model from A to B. You practically built a wall here and a wall there so the thing would go in a straight line. Models are now very steerable. A lot of that steering you can just put in the prompt: go from A to B, and it will. If you have harnesses designed to do that steering, you can delete that part. They frequently encourage it. That is what people mean when they say the model will consume some of the scaffolding. What the harness needs to start doing is let the model run longer. You do not want it to stop at B. You want B to C to F to Z and then back to A. That sounds slightly silly, she says, and it produces a lot of differences: less steering harness, more strategy harness, matching the intelligence gains in the models.

Do task-specific or vertical-specific harnesses make sense? Their opinion is yes. There is no general harness. Coding is a general capability because software has eaten so much of what is possible. Very specific domains will need pieces customized. One of those is how you handle errors between doing something and handing it to the model. In domains that need extreme verification — legal and finance, where getting it wrong has consequences — that last bit of logic gives you a ton of juice. It is the difference between your product and someone else's. Context, Jiang thinks, is a little overdone. Any harness can handle a lot of context; if you have the data, you are uniquely qualified. Lesse adds that "harness" means a lot of different things, which is why the opinions collide: a user-model-tool loop to some people, a packaged toolset to others. The generic bits — caching, clearing old tool calls — are less interesting to own. Go a layer higher and you start wanting control.

Huang names the moment: the industry went through token maxing, and now it is in token rationalization. Both guests take the frame. Jiang says it makes sense. As models get more capable you hit levels of intelligence maxing, then the next dimension is cost or speed, and you walk that across the distribution of task complexity. What you do not want to do is stop AI usage. They see customers do it. Spend often erupted as shadow IT: employees found a way, procured it themselves, and before you know it half the org has installed Claude Code. The wrong move is a hard cap. If you are getting returns — shipping faster, running more operationally efficient — those are gains. The better architecture is a strategy that assesses a task's complexity and routes: hard work to a big, smart model, easier work to cheaper ones. There is technical complexity in it. It is very doable.

Will they offer that routing? Within the Claude space, Jiang thinks it will make sense. It is one of the strategies they imagine designing. Every month feels like a new era; they want recomposable primitives so they can redesign quickly. On model routing they feel strongly about designing the platform for Claude. They will restrict to that space. They are not that interested in telling you to route to a different vendor's model. Lesse's supporting belief is that harnesses and the agentic layer should be tuned to the model family you use. There was a period when people thought they could build a harness and an agent and plug in a different model underneath, and they were excited about routers for that reason. Players have started coming up a layer. Vercel, she notes, did this with its harness-and-agent offering: plug in the whole thing tied to a model family. What Anthropic can provide is smarter mix-and-match among models inside that family.

Before Anthropic, Lesse was at Stripe, in the very reasonable era of watching the AWS bill. If someone built a background job, misconfigured it, and burned CPU on something not worth the spend, the guardrails found it and someone asked that engineer, nicely, to please turn it off. AI spend is heading for the same muscle. It gets dangerous when you just set a cap. The smarter move is to come in from the side afterward: there are a few ways you could have accomplished that outcome. One is take Opus and run it all night. Another is get a little smarter with the strategies and hit the same outcome at lower cost. That, she thinks, is the next layer of thinking. Usage mix is shifting too. Coding exploded, of course. Manufacturing has started to pick up; one of their PMs got on a flight to Detroit to see what those customers actually need. They expect more outside-the-box categories.

Asked what they are excited to build in the next few months, Jiang apologizes for saying the word again and says it anyway: composing strategies, moving into that coordination layer. The problems people are building sit at a layer where getting return takes cleverness about the nature of the problem. Concrete case: a bug-hunting agent. You can send one off and get a certain return. People get stuck. The next options look like swapping in a bigger model or letting it run longer. From a lot of experimentation those two levers are still true, and there is a third that tends to do more than you think: best-of-N. Saying the words is fine; there are papers. Building it, putting it in production, testing it on users — that is really hard, and you end up building a pile of custom harnesses. That is where the alpha is. If it gets you the return you want and it is hard, their opening philosophy still holds: try to make it easy so you can run the experiments you actually need.

Reeder hears last year's agent swarms. Jiang says that is a type of strategy, the way one big agent that separates into several is another. People have thought about it as human organization. Taken to the end state, it is more that the token has a job. That job piece is what they are indexed on. Internally they can give you five jobs off the top of their heads. Give this out to the rest of the ecosystem and it is probably 100,000 or 200,000 combinations. Who knows. Lesse wants to keep hill-climbing intelligence per dollar and put that power in people's hands. Around the edges are two personas. Enterprises need security and compliance, and they need the platform more modular in the right ways — plug memory into this thing over here — plus a truly excellent developer experience, because many of them live in a walled garden and cannot use the clever strategy work for XYZ reasons. Solving that is important. The other persona is the weekend developer building something useful for themselves on Anthropic and on everyone else's platforms. For them, more open, more hackable, go-wild solutions. A lot of that is table stakes, and she is excited about it, because table stakes are what let people then plug into the hill-climbing.

Pull the hour into a list you can use tomorrow and it looks like this. Turn on prompt caching today; clear old tool calls out of the window. Audit the harness and delete the steering scaffold; put the saved attention into evals. Leave some of the token budget for reflection, memory, advisers, and graders. When you stall on returns, try best-of-N before you only reach for a bigger model. If you are in legal or finance, hold the verification logic between the model and the execution system — that is the moat. If you are an enterprise staring at the bill, do not slam a cap on; build complexity routing, then review the jobs the way you once reviewed AWS. If the system has no API, computer use can run now. Jiang's electricity image still fits at the end: the candle era could only do so much. The value of the current is how many things you can plug in. Competition at the platform layer is about sockets, and about getting the current into more of the business.