Decoupling the Brain from the Hands: Claude Managed Agents (Part 1 of 4) — Anthropic
Source
youtube.com
Author
AI Engineer
Date
Context
Anthropic redesigned Claude Managed Agents so a failed execution environment need not erase the AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition →'s working history. Its original design put the event log, agentic loopThe cycle an agent runs in: decide, call a tool, read the result, decide again — repeating until the goal is met or a stop condition fires.Full definition → and code execution in one container. The replacement separates them: a durable log records events, the agent loop decides what to do, and replaceable sandboxes execute code. A replacement agent loop can resume from the log; a sandboxAn isolated environment where AI-generated code or agent actions run without being able to touch anything real.Full definition → failure becomes a tool error the agent can handle. The workshop develops the recovery design described in the engineering write-up, also covered in earlier Intel.
Key takeaways · AI-distilled
Phatak defines an agent as a model in a loop with tools plus a agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition → and a sandbox, and argues harnesses encode model-specific assumptions that go stale with each new model release.
The first Managed Agents design kept session, harness and sandbox in one 'pet' container that was hard to debug; the redesign splits an append-only session log, a brain (harness plus Claude) and replaceable hands (sandboxes and tools).
Credentials are injected at an egress proxy so Claude never sees them inside the sandbox, and Phatak reports time to first token dropped about 90% after the split.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agentic loop — The cycle an agent runs in: decide, call a tool, read the result, decide again — repeating until the goal is met or a stop condition fires.
sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
video1,000 Agent Tasks in a Sandbox: What Breaks When LLMs Run Code — Kevin Orellana, AWS
videoHow Inference Keeps Up with Our Insatiable Need for Tokens — Cerebras & Gimlet Labs
videoAuth for Agents: Let AI Sign Up for Your Product with auth.md — Michael Grinich, WorkOS
videoThe Next Trillion Internet Users Still Don’t Have an Identity — Adi Singh, AgentMail
