← back
Claude for Long-Horizon Tasks — Lance Martin, Anthropic
Takeaway
Long-running agents need durable sessions, replaceable execution environments, and independent outcome verification.
Summary
- Longer model task horizons enable asynchronous agents and motivate APIs that package both an agent harness and deployment infrastructure.
- Claude Managed Agents separates a stateless harness from execution containers, persists sessions in an append-only event log, and keeps credentials in a separate vault.
- Persistent session history lets agents retrieve older context after compaction rather than irreversibly discarding it.
- Independent verifier contexts evaluate a builder's output against measurable goals and drive repeated improvement until the desired outcome is confirmed.
long-horizon-agentsclaudeverification
Original description
Claude is capable of long horizon tasks. In this talk, we'll share lessons learned about building agent harnesses for reliable and secure long-horizon work. This include decoupling the brain and hands, self-verification, self-learning, and design for evolving agent harnesses. Lance Martin Member of Technical Staff · Anthropic [X/Twitter](https://x.com/RLanceMartin) · [LinkedIn]( / lance-martin-64a33b5 ) · [Website](https://rlancemartin.github.io) Member of technical staff at Anthropic. Working on the Claude Platform, including Claude Managed Agents and the claude-api skill in Claude Code. Prior to Anthropic, was one of the early team at LangChain. Prior to LangChain, spent several years focused on vision for self-driving cars (Uber ATG, Ike, Nuro) and got a PhD from Stanford.