Claude Managed Agents
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Claude Managed Agents is an API released by Anthropic that packages both the agent harness and managed deployment infrastructure, providing a structured approach to building and running long-horizon AI agents on top of Claude.
Overview and Architecture
Lance Martin describes Claude Managed Agents as combining two components that are otherwise left to developers to wire together: the harness (the "brain") and the execution environments (the "hands"). This separation is explicit by design — Martin notes that the system is built around "decoupling what we call the brain, that's the harness, from the hands, the execution environments." →
The API also abstracts away managed deployment infrastructure, meaning developers do not need to build or maintain the scaffolding typically required for running persistent agent sessions. →
Session State Design
A core architectural decision in Claude Managed Agents is how session state is represented: the session is structured as an append-only event log. → This design choice has implications for auditability and reliability in long-horizon tasks, as state accumulates sequentially rather than being mutated in place.
Practical Usage
Martin demonstrates applying Claude Managed Agents to agent verification workflows, describing setting up "a kind of a verifier loop using Managed Agents and outcomes" → — illustrating that the API is suited not only for primary task execution but also for implementing oversight and correctness-checking loops within an agentic pipeline.