anthropic / claude-managed-agents / linear

Anthropic’s Claude Code product lead outlined the startup’s approach to usage limits, transparency, and what it calls a “lean harness” for safety and cost control. The discussion covered per-user usage caps, clear signaling to developers about model behavior and constraints, and lightweight safety layers that aim to balance developer freedom with guardrails. The lead emphasized transparency about limitations and predictable costs as key for developer trust and broader adoption, noting trade-offs between model capability, safety, and infrastructure cost. This matters because it shows how an AI company operationalizes safety and monetization strategies for developer platforms amid competition with major AI providers.

Anthropic’s claude-cookbooks repository added a new managed_agents/linear example that bridges Linear issue events to Claude Managed Agents. The commit introduces environment examples, gitignore entries, and a CLAUDE.md guide explaining a stateless webhook flow: Linear AgentSessionEvent → CMA session → session.status_idled webhook → createAgentActivity reply. It provides setup hints (invoke /claude-api, follow skill.md) and outlines extensibility: mounting GitHub repos, enabling MCP tools, outcomes, multiagent configs, memory stores, and custom tools. The changes give developers a practical template and checklist for integrating Linear with Claude agents, reducing friction for building automated agent workflows and real-world tool integrations.