Quartz 5

Launching Claude Managed Agents

Properties1
tagsclaude-code, llm, agentic-workflow, agent

3 min read

Summary

Claude Managed Agents is Anthropic’s pre-built agent harness running on managed infrastructure, designed to address two problems: agent harnesses encoding stale assumptions about Claude’s capabilities, and growing task horizons requiring resilient, long-running infrastructure. The system introduces three core primitives — Agent (versioned config with model/prompt/tools/skills), Environment (sandbox template), and Session (stateful execution) — and decouples the “brain” (Claude + harness) from “hands” (sandboxes/tools) and “session” (event log) so each can evolve independently.

Claude Managed Agents 是 Anthropic 推出的預建代理人框架，在受管基礎設施上運行，解決兩個問題：代理人框架對模型能力假設過時，以及任務時間範圍增長對基礎設施的需求。引入 Agent、Environment、Session 三個核心原語，並將大腦、雙手與會話解耦。

Key Points

  • Core primitives: Agent (versioned config: model, system prompt, tools, skills, MCP servers), Environment (sandbox template: runtime type, networking, package config), Session (stateful run: provisions sandbox, mounts resources, stores auth in vault)
  • Problem solved: Claude’s task horizon already exceeds 10 human-hours on METR benchmark; conventional harnesses can’t keep pace with capability growth
  • Common patterns: Event-triggered (bug flagged → agent writes patch → opens PR), Scheduled (daily briefs of X/GitHub activity), Fire-and-forget (assign via Slack, get deliverable back), Long-horizon tasks
  • Architecture principle: decouple brain from hands from session so each can fail or be replaced independently — this gives reliability, security, and flexibility
  • Onboarding: claude update then /claude-api managed-agents-onboarding skill
  • Six SDK languages: Python, TypeScript, Java, Go, Ruby, PHP

Insights

The architectural decoupling of brain/hands/session is the key infrastructure insight — most agent failures compound because a harness bug, a sandbox crash, and a session corruption all look the same and are hard to isolate. The fact that Anthropic itself is moving from the Agent SDK to Managed Agents as the recommended primitive signals this is the direction for production workloads. The “harness encodes assumptions about what Claude can’t do” problem is a concrete manifestation of the bitter lesson: hand-coded constraints become technical debt as model capabilities improve.

Connections

Raw Excerpt

Building agents to scale with Claude’s intelligence is an infrastructure challenge, not strictly a matter of harness design.

Graph View

Backlinks