How Anthropic ships Claude with Claude: Separating the Brain from the Hands — Latent Space
Source
latent.space
Author
Spooky Space 👻 AI News, Views, & Ghosts
Date
Why it matters
Explains why preserved thinking blocks can break a custom harness, how credentials stay outside the sandbox, and how Managed Agents separates harness, session and sandbox, all useful when designing agent runtimes.
Key takeaways · AI-distilled
To resist chain-of-thoughtHaving a model write out intermediate reasoning steps before its answer, which markedly improves performance on hard problems.Full definition → extraction for distillationTraining a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.Full definition →, Anthropic has enforced preserved thinking blocks across turns over its last couple of models. That broke some customers' manual context compactionSummarizing an agent's earlier conversation to free room in the context window so a long session can keep going.Full definition →, so it shipped an API compaction feature that works with preserved thinking.
The guests describe porting the Python-monolith Messages API to Rust with no Rust developers on the team: Claude wrote the code, humans designed the system, and shadow traffic checked outputs matched. They estimate roughly 40-50% CPU savings and 10-20% faster time to first token.
In Claude Managed Agents, credentials sit in a separate vault and the agent sees only placeholders; an egress proxy injects the real credential on the way out. Sandboxes also support domain allowlists or no network, so untrusted generated code cannot reach arbitrary hosts.
The guests say many publicized sandbox escapes happen in training, where models are boxed with impossible tasks and pushed to break out, not in production, where online safeguards run and the harness sits apart from the execution sandbox.
For multi-agent work on Opus 5.5, the guests advise one lead agent that splits parallelizable tasks, with sub-agents that share uniform context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → and run asynchronously without blocking each other. Beyond those two properties, they suggest letting Claude design the architecture.
Terms in this piece · Glossary
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
chain-of-thought — Having a model write out intermediate reasoning steps before its answer, which markedly improves performance on hard problems.
context compaction — Summarizing an agent's earlier conversation to free room in the context window so a long session can keep going.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
