Anthropic engineers detail how Claude’s ‘brain’ and ‘hands’ are separated for safety and performance

Katelyn Lesse and Angela Jiang explain the managed-agent architecture behind Opus 5.5 and the company’s platform strategy

By LineZotpaper AI staff

Published27 October 2026

Read Time2 min

Sources2 outlets

Follow this storyHear when it moves

Illustration: Anthropic engineers detail how Claude’s ‘brain’ and ‘hands’ are separated for safety and performance

Anthropic’s engineering and product leads have described how the company ships its flagship Claude models by separating the model’s reasoning, the ‘brain’, from the execution environment, the ‘hands’, a design choice they say improves both security and latency. The insights come from a podcast interview with Katelyn Lesse and Angela Jiang, heads of Engineering and Product respectively, as Anthropic prepares for its expected IPO.

The architecture, known as Managed Agents, decouples the session log and harness from the sandbox that runs code. The harness persists while execution environments are created on demand, meaning a failed sandbox does not destroy the entire agent session. The same separation allows credentials to be held outside the sandbox and attached through a proxy, so generated code never has access to service credentials. Network controls further restrict where sandbox traffic can go.

Lesse and Jiang said the earlier approach placed the whole coding agent inside a sandbox behind an API, adding latency before the first response and risking the loss of the running agent if the sandbox disappeared. Splitting the components lets Anthropic pursue lower latency and better reliability while keeping execution isolated.

The discussion also covered the launch of Opus 5.5, which the sources describe as the current #1 LLM by general consensus. Anthropic says the new model costs less to run than its predecessor, but the exact saving depends on workload and cached-context usage. The team addressed feedback that earlier Opus models were too verbose, and worked to improve intelligence per dollar.

On the safety side, the team noted that as models become more capable, safe deployment decisions grow harder. The company protects model reasoning against extraction used for distillation by preserving thinking blocks across turns, though that can break some customers' existing harnesses. Lesse described early-access programs as a process where customers run evaluations in their own environments and report surprising behaviour, seeing the model in tasks Anthropic’s own products do not cover.

The podcast also touched on platform changes such as rebuilding the Messages API in Rust, and on concepts like “dreaming,” in which one agent inspects another agent’s prior work and writes memories back, built on the same managed-agent primitives. The team said its immediate focus has shifted toward safety as model launches have become more demanding, while keeping the base harness simple so developers can experiment with verifiers and parallel agents on top.

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Analysis

Why This Matters

The architecture described by Anthropic offers a blueprint for building secure, low-latency agent systems, directly affecting the many developers integrating model APIs into their products.
Anthropic’s approach to separating credentials from execution could influence broader industry practices around AI safety and containment.
As the company approaches its IPO, the efficiency gains from this design, alongside its turnaround to profitability, signal how frontier AI labs can build sustainable businesses.

Background

Anthropic is a US-based AI research and deployment company, best known for its Claude family of large language models. The company has positioned safety as a core value, and its platform provides API access to developers. The Managed Agents architecture is an internal system that powers Claude Code and other agentic features. The discussion took place as Anthropic filed its S-1 ahead of a potential listing.

What to Watch

Whether other labs adopt similar brain/hands separation in their agent platforms.
How Anthropic’s safety classifiers and model-routing systems evolve as frontier models are released more frequently.
The impact of the Messages API rewrite in Rust on API performance and developer experience.

Sources

How Anthropic ships Claude with Claude: Separating the Brain from the Hands · Latent.Space
How Anthropic ships the #1 LLM in the World safely: Separating the Brain from the Hands · Latent.Space

Checked

12 claims in this story checked against 1 source: 11 supported.

Show each claim

Partly supportedThe headline says Anthropic engineers described the separation, although one of the two speakers is identified as a product lead, not an engineering lead.Latent.Space: “Katelyn Lesse and Angela Jiang, heads of Engineering and Product respectively.”
SupportedKatelyn Lesse and Angela Jiang explained the managed-agent architecture and Anthropic’s platform strategy.Latent.Space: “we dive into its launch, and how Anthropic ships Claude with Claude with Katelyn Lesse and Angela Jiang, heads of Engineering and Product respectively.”
SupportedThe source describes Opus 5.5 as the generally accepted number-one LLM.Latent.Space: “it has lasted a month as the generally accepted #1 LLM in the world”
SupportedAnthropic’s Managed Agents separate the session log and harness from the code-executing sandbox; the harness can persist if an execution environment fails.Latent.Space: “In Managed Agents, Anthropic separates the session log and harness from the sandbox that executes code. The harness can persist while an execution environment is created when needed; a failed sandbox does not have to take the whole agent session with it.”
SupportedThe separation is described as supporting security and lower latency while keeping execution isolated.Latent.Space: “Separating the pieces allowed the platform to pursue lower latency and better reliability while keeping execution isolated.”
SupportedCredentials are held separately and attached through a proxy, so generated code in the sandbox cannot read service credentials.Latent.Space: “The episode describes credentials held separately and attached through a proxy, alongside controls on where sandbox traffic can go.”
SupportedThe earlier approach put the whole coding agent in a sandbox behind an API, adding work before the first response and risking the loss of the running agent if the sandbox disappeared.Latent.Space: “The guests say an earlier approach put the whole coding agent in a sandbox behind an API. Starting the environment and agent added work before the first response; if the sandbox disappeared, so did the running agent.”
SupportedAnthropic says Opus 5.5 costs less to run than its predecessor, with the saving depending on workload and cached-context use.Latent.Space: “Anthropic says the new model costs less to run than its predecessor; the exact saving depends on the workload and its use of cached context.”
SupportedThe team worked to address feedback that earlier Opus models were too verbose.Latent.Space: “We discuss feedback that the previous Opus generations were be too verbose, alongside the need for stronger intelligence per dollar.”
SupportedPreserving thinking blocks across turns helps protect model reasoning from distillation, but the requirement can break customers’ existing harnesses.Latent.Space: “Preserving thinking blocks across turns helps address this, but a new requirement can also break a customer's existing harness.”
SupportedLesse described early-access customers running evaluations in their own harnesses, reporting surprising behaviour, and showing Anthropic tasks its own products do not cover.Latent.Space: “Customers run evaluations in their own harnesses, report surprising behavior, and sometimes pass along less formal impressions. The useful part is not just a benchmark score: it is seeing the model in environments and tasks Anthropic's own products do not cover.”
SupportedThe team said more of its immediate attention had shifted to safety as model launches became harder.Latent.Space: “Memory and learning remain important, but the team says more of its immediate attention has moved to safety as model launches have become harder.”