Unverified50% confidenceFactExact time
The refactored Managed Agents is split into three interfaces: an append-only session log (Session), the Agent loop of Harness plus Claude (Brain), and sandboxes and tool calls provisioned on demand and destroyed after the task ends (Hands).
1
Sources
50%
Confidence
Long-term
Relevance
10/25/2026
First Seen
Sources
How Anthropic Decoupled the Brain and Hands of Claude Agents
bilibili量子菠萝_10/24/2026
Related Entities
Related Claims
UnverifiedAn Agent consists of three parts: the LLM (brain), tools (hands), and memory (diary), and none of the three can be missing.74% similarUnverifiedA durable and reliable Agent system needs a three-piece kernel: memory storage, tool contracts, and guardrails; everything else (planning loops, reasoning chains, meta-prompting) is upper-layer scaffolding that can be replaced independently.70% similarUnverifiedCompared to single-turn dialogue, Agent systems introduce three additional dimensions: memory, tool invocation, and planning.69% similarUnverified复杂任务的 Agent 执行拆成三个角色：调度智能体做研究规划、执行子 Agent 在隔离 WorkTree 里并行工作、Verifier 跑测试检查结果并将证据附到 PR68% similarUnverifiedIn the agent context, a single turn may contain multiple tool calls, and the SDK executes them sequentially and appends each step's result to the context until Claude decides to stop calling tools.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1095491API
curl https://kongchang.com/api/v1/knowledge/claims/1095491MCP
get_claim(id=1095491)