Type: url
Original: https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
Status: generated
Summary
An engineering post on getting an agent to make consistent progress across many context windows, where each session begins with no memory of the last. It reports two observed failure modes — the agent trying to one-shot the whole app and running out of context mid-feature, and a later session seeing existing progress and declaring the job done — and describes a two-part answer: an initializer agent that sets up the environment once (an init.sh script, a progress log, an initial commit, and a JSON feature list of every requirement marked failing), and a coding agent that each session reads that state, advances one feature, verifies it end-to-end with browser automation, and leaves the repository clean and committed.