One Unit of Work per Session

Limiting each session or loop iteration of a long-running agent to one feature or item prevents one-shot attempts and premature "done" claims.

Evidence: moderate

Given a large goal and an open-ended session, agents tend to attempt everything at once, exhaust their context half way through, and leave partial work that a later session mistakes for completion. Anthropic names both of these, one-shotting and premature completion, as the failure modes its long-running harness was built to stop.

Its fix is to scope each session to a single feature. The coding agent picks one unfinished item, implements it, commits with a descriptive message, updates the progress file and leaves the code in a clean state fit to merge. Geoffrey Huntley's Ralph loop reaches the same rule from the other direction. It is a plain shell loop that pipes the same prompt file into a coding agent over and over, loading the same specs and plan each time and taking one item per iteration. What makes such a crude loop converge is back-pressure: type checkers, linters and tests give a correctness signal on every pass. Huntley also allows many sub-agents for search but only one for build and test, the rule of single-threaded writes, and tunes the loop by adding constraining prompt lines where a failure is predictable.

The evidence is two practitioner-level accounts, one from a lab and one from an individual, with no comparative numbers.

The rule fits the failures that grow with time in one context, such as stopping early and losing constraints to compaction; see long-context agent failure modes. Each session pairs it with a start-of-session smoke test on entry, verification before done on exit, and externalised progress artefacts to carry the outcome to the next session. Outside coding, the unit is whatever the agent can finish and verify in one context: one incident, one ticket, one document section.

Sources

  • Anthropic Engineering (Justin Young), 26 Nov 2025 — https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
  • Geoffrey Huntley, undated — https://ghuntley.com/ralph/