Summary
Engineer at Anthropic. Authored “Effective Harnesses for Long-Running Agents” (2025-11-26), addressing how coding agents maintain progress across multiple context windows.
Key Points
- Identified two failure modes for long-running agents: over-ambition (context exhaustion mid-task) and premature completion (sees progress, declares done)
- Proposed Initializer + Coding Agent two-phase pattern with file-based handoff
- Key insight: use feature list JSON as a contract — agent only modifies
passesfield, can’t delete requirements - Advocates one-feature-at-a-time + git commit for incremental, recoverable progress
- Framing: agent session handoff mirrors human team handoff mechanisms
Open Questions
None currently
Evidence Timeline
- 2026-04-07: Created from “Effective Harnesses for Long-Running Agents” article (published 2025-11-26)