AI Research Atlas

Effective harnesses for long-running agents

Anthropic · 26 November 2025

Anthropic's fix for agents that lose track across context windows is an initializer agent plus a coding agent that leaves artifacts for the next session.

Compaction alone was not enough, since even Opus 4.5 running in a loop on the Agent SDK across context windows fell short of a production-quality app. The two-agent harness sets up the environment once, then makes incremental progress each session and documents state, like engineers handing over shifts.

Date
Wednesday, 26 November 2025
Lab
Anthropic
Kind
paper
Access
paper only

Qualitative engineering guidance with a code quickstart; no benchmark was read. Continued in 'Harness design for long-running application development' (2026-03-24).

Sources

  1. www.anthropic.com/engineering/effective-harnesses-for-long-running-agents

This record was checked against its sources on 6 October 2026. How we check

Related