How Anthropic ships Claude with Claude: Separating the Brain from the Hands

What it claims

This is a Latent Space interview with Katelyn Lesse (head of engineering) and Angela Jiang (head of product) for the Claude Platform at Anthropic, recorded the day after the Claude Opus 5.5 launch. Its main point for API users is that a model launch now carries integration changes. Opus 5.5 enforces preserved thinking blocks across turns, including across multiple accounts, to block chain-of-thought extraction for distillation. Anthropic shipped a server-side compaction feature for harnesses whose manual compaction would otherwise break under this rule. Lesse says safety and access work is now "the majority of the work" in a launch.

How it works

Distillation defence. Anthropic sorts user-experience costs into two tiers. One is an annoying integration change. The other is losing a capability. Anthropic says it accepts the first and tries hard to avoid the second. Before launch it contacted open-source harness maintainers and large API customers about the required upgrades. When preserved thinking first appeared with Fable 5.1, existing customers got a grace period, which was enforced through a trust grant.

Trust grants. This is Anthropic's internal authorization data model. A grant gives an account access to a specific model, a verification program, or an exemption such as that grace period. A service on the request hot path checks it. Lesse calls it "really not any different than vanilla authorization." The same primitive supports the cyber and life-science verification programs (which relax online safeguards for approved, scoped use cases), access to Mythos, and a single control that can withdraw model access from everyone.

Per-model retention policy. Each launch gets its own data-retention decision. Opus 5.5 has no data-retention requirement. Fable does. Retention exists so automated and human review can detect abuse spread across many accounts, which online classifiers cannot catch.

Brain and hands. Managed Agents keeps the session log and harness apart from the sandbox that executes code. An earlier design put the whole coding agent in a sandbox behind an API. That added startup work before the first response, and a lost sandbox took the running agent with it. Credentials never enter the sandbox. A proxy attaches them, and the sandbox's network egress is controlled.

flowchart LR
  H[Harness + session log<br/>persists] -->|creates on demand| S[Sandbox<br/>runs generated code]
  S -->|traffic| P[Proxy<br/>egress controls]
  C[Credential store] --> P
  P --> E[External services]
  S -. sandbox failure .-> H

Managed Agents and Anthropic's first-party products share the sandbox fleet and API infrastructure but not the harness. Anthropic keeps the harness intentionally simple and expects developers to add verifiers, parallel agents and similar strategies on top. "Dreaming" is one agent reviewing another agent's past work and writing memories back.

How to use it

• Harnesses that replay, strip or rewrite thinking blocks need changes. Replace manual compaction with the API compaction feature.
• Claude Code's default reasoning effort is now medium. Defaults vary by segment (free, prosumer, Claude Code, Cowork, Enterprise, API), so set effort explicitly if behaviour matters to you.
• The guests say Managed Agents exists partly because the mix of API parameters (effort, budgets and others) is hard to tune. Anthropic retunes its harness for each new model.

Limits

The page is truncated at about 60,000 of 129,786 characters. The Rust rewrite of the Messages API (56:05), CPU constraints, the credential and sandbox discussion (1:05:25), and safety classifiers and model routing are known only from timestamps and the editorial intro. The transcript is an uncorrected automatic draft. The page gives no API field names, parameters or exact pricing. Early access goes only to customers with robust eval suites, and the guests describe feedback as partly "vibes-based." The editors say future platform ideas are plans, not commitments. The host's S-1 figures conflict: the intro says a -126% loss, while the transcript says losing 129% of each revenue dollar a year earlier.