Anthropic reveals how it builds Claude with Claude: the brain separate from the hands

Anthropic has filed its S-1, and on the same day Latent Space released a conversation with Katelyn Lesse (engineering) and Angela Jiang (product) about how the company ships Claude using Claude. The backdrop is telling. According to figures cited by the host, 12-month revenue per employee grew from $1.4M to $6.6M, nearly a fivefold increase. The company went from a loss of about −126% of revenue to a 19% GAAP profit. The host says the IPO is expected in roughly two weeks, with a valuation of $2 trillion.

According to the S-1, two loops drive this efficiency. The improvement loop: Claude makes models better, better models open up more use cases, and those use cases generate feedback. The investment loop: revenue from growing adoption is turned into compute and new capabilities. The conversation shows what this looks like at the API and agent-runtime level.

Brain and hands: how the agent was split in two

The episode's central architectural idea concerns Managed Agents. Anthropic separated the session log and harness from the sandbox where code runs. The harness is long-lived, while the execution environment is created on demand. If the sandbox crashes, the agent session continues.

The old design was different: the entire coding agent lived inside a sandbox behind the API. That caused two problems:

• before the first response, the system had to spin up both the environment and the agent itself, which added latency;
• when the sandbox disappeared, the running agent disappeared with it.

The split had a side effect for security. Generated code runs in a sandbox that service credentials never enter, so there is nothing there to read. Credentials are stored separately and injected through a proxy, and the sandbox's network traffic is restricted by direction. For Anthropic, this is a matter of both protection and performance.

Launching a model like launching an API

Opus 5.5, which the host considers the best LLM in the world for the past month, was built with a focus on quality and cost. Earlier Opus models were criticized for verbosity, so intelligence per dollar mattered. Anthropic says the new model is cheaper to run than its predecessor. The exact savings depend on the workload and the share of cached context.

The downside affects integrations. To make it harder to extract reasoning for distillation, thinking blocks must be preserved between conversation turns. But the new requirement can break a customer's existing harness. That is why early-access programs matter: customers run the model in their own harnesses and find behavior that doesn't show up in Anthropic's products.

Other topics from the episode:

• Dreaming. The implementation is simple: one agent studies another's earlier work and writes memories back. It is an example of high-level behavior assembled from the same Managed Agents primitives.
• Priorities. Memory and learning remain important, but the team's attention is shifting toward safety, because releasing models has become more complicated. Decisions about who gets access to models have become a task in their own right.
• Base harness. The team tries to keep it simple. Developers build verifiers, parallel agents, and other strategies on top. Anything mentioned as a future platform capability remains a plan.
• Infrastructure. The episode's agenda includes CPU limits when scaling and the rewrite of the Messages API in Rust. At this rate of growth, even the CPU cost of a single request becomes a strategic choice.

Claude Platform as the foundation for everything

The S-1 describes Claude Platform as a comprehensive environment for building agents. More than 50% of customers with annualized revenue of $100K or more use it. Claude Code and Cowork are built on the same primitives that are open to external developers. Anthropic now looks not only like a lab but also like a platform whose role is close to that of a hyperscaler. In its IPO narrative, the company calls agents "virtual collaborators": they own a goal and work alongside a human rather than inside a single application.

For anyone building their own agent runtime, the practical takeaways come down to three questions:

1. Which parts of the system must survive a failure? Move the log and harness out of the execution environment. 2. Where does untrusted code run? In an isolated sandbox with a restricted network. 3. Which secrets really need to be where the code executes? Most often none: a proxy is enough.

It is also worth investing separately in evals. A harness can be treated as an application, but improving it looks more like research: goal, measurement, error analysis, iteration. Models help build and analyze evaluations, but a human must decide whether a test reflects real user work. If your harness breaks because thinking blocks are preserved, it is better to find out during early access than on release day.