How Anthropic’s Managed Agents Cut AI Agent Latency by 90% with Brain‑Hand Decoupling

Why AI Agents Are Hard

AI agents must call code‑execution tools, run for minutes or hours without dropping, retain context across steps, stay isolated from production systems, and recover from step failures—tasks that previously required months of engineering effort.

What Managed Agents Provide

Anthropic bundles four core capabilities:

Production‑grade infrastructure – sandbox isolation, authentication, and tool calling are fully managed.

Long‑running sessions – agents can run for hours without losing progress, e.g., processing hundreds of contracts.

Multi‑Agent collaboration (research preview) – one agent can launch others to handle complex workflows in parallel.

Trusted governance – fine‑grained permissions, identity management, and end‑to‑end tracing ensure safe AI access to real systems.

Internal tests on structured‑file generation show up to a 10‑point increase in task success rate, especially on harder tasks.

Design of the System

Initial mistake: packing the Claude runtime, code‑execution environment, and session logs into a single container, turning it into a "pet" that crashes and loses all in‑flight work. This also created a security risk because generated code and API keys shared the same environment.

Solution: Separate Brain and Hand

The core decision is to fully decouple three components:

Brain (Claude + Harness) – performs reasoning and issues commands.

Hand (Sandbox) – executes code and manipulates files.

Session logs – record all events in independent storage.

If the hand crashes, the brain receives an error and can retry by spawning a new hand. If the brain crashes, a new brain reads progress from the session log and continues. Logs are stored separately, so any component failure leaves the log intact for recovery.

Performance Impact

Decoupling yields a surprising speed boost: p50 latency drops ~60% and p95 latency drops >90% because the Claude model starts reasoning immediately and only launches a sandbox when code execution is needed.

p50 is the median waiting time; p95 is the 95th percentile. A >90% drop in p95 means the worst‑case delays that used to last minutes are now almost gone.

Key Security Improvement

Credentials are kept outside the execution sandbox in a vault and accessed via a dedicated proxy, so the sandbox never sees raw tokens.

Git repositories – after cloning with an access token, the token disappears; Claude uses the local config instead.

Third‑party tool (MCP) – Claude sends a request to the proxy, which fetches credentials from the vault and forwards the call, keeping the key hidden from Claude.

Real‑World Deployments

Notion – assign tasks to Claude inside Notion, run multiple agents in parallel, complete everything from code to PPT within Notion.

Asana – AI teammates take tasks in Asana projects, dramatically speeding up development and shifting engineering effort to UX.

Rakuten – five professional agents (product, sales, marketing, finance, HR) integrated with Slack/Teams, launched in under a week.

Sentry – bug root‑cause analysis → auto‑generated fix → PR, reducing weeks‑long cycles to a few weeks.

Atlassian – assign tasks directly in Jira workflows, weeks to launch without building sandbox or permission systems.

Vibecode – turn a prompt into a deployed app, achieving >10× faster setup.

All six companies compressed infrastructure setup from months to days or weeks.

Takeaways for Building Your Own AI Agent Platform

Decouple components so each can fail and recover independently.

Externalize state (session logs) to enable restart from the last known point.

Never let credentials enter the execution environment; use proxies or vaults.

Prioritize stable interfaces over specific implementations, e.g., execute(name, input) → string, so underlying containers or APIs can change without breaking agents.

Pricing and Access

Managed Agents are in public beta, billed at $0.08 per session‑hour plus standard Claude API token fees. Access points include the Claude Platform, documentation, and a quick‑start console. Multi‑Agent collaboration and self‑evaluation are currently research preview features requiring separate access.