Anthropic engineers describe splitting Claude agents into brain and hands

An agent that holds credentials in the same place it runs untrusted code is one bug away from a breach. Anthropic now says the fix is to split those jobs apart. Latent Space published an interview on 27 October. The guests are Katelyn Lesse and Angela Jiang, Anthropic's heads of engineering and product. They describe how the company builds and runs Claude at scale.

Lesse calls the design "brain and hands". The brain is the agent harness. The hands are the sandbox that runs code. In her own words, "the harness runs on a durable server, really ours, right? And the hands are like an execution environment, like a sandbox or a container, right?"

Why the two halves live apart

The agent and the credentials never share a process. Lesse told the podcast: "the agent is never like in the same place where the credentials are... there's a literal infrastructure wall that says the agents like can't access those credentials."

The split also helps speed. A durable harness holds the session state. A sandbox or container starts only when the agent needs to run code. In Lesse's words, that means "we can get you tokens back really quickly". A failed sandbox does not take the whole session down with it.

A Rust rewrite with measured gains

Lesse also described a rewrite of Anthropic's Messages API. The team moved it from an older stack to Rust. The result, in her own words, is "somewhere like 40 to 50 percent" CPU savings. She also said the team saw time-to-first-token improve by "somewhere between like 10 and 20 percent". The caveats are hers.

The numbers are Anthropic's own. They come from one speaker on one podcast. No benchmark is published. No third party has repeated the measurement. But the shape of the claim is specific. It names the service rewritten. It names the language it moved to. It gives two measured outcomes.

The podcast is tied to Anthropic's recent S-1 filing. Host Shawn Wang gives two numbers in his opening summary. Trailing-twelve-month revenue per employee rose from 1.4 million dollars to 6.6 million dollars. The company moved from a 129 percent GAAP loss to a 19 percent GAAP profit over the year. The two figures are the host's own summary of the S-1. They are not direct quotes from Lesse or Jiang.

A team building its own agents can copy the brain and hands split. Keep the long-lived harness on a trusted server. Give it a short-lived sandbox for every piece of code the agent writes. Keep credentials and user data out of the sandbox. The design is simple to describe. It is harder to run. It is the design one of the busiest agent services on the internet now runs on.

Source

How Anthropic ships Claude with Claude: Separating the Brain from the Hands, Latent Space