Externalised Progress Artefacts

Long-running agents keep continuity across context windows through artefacts outside the transcript (a pass/fail feature list, a progress file, an init script, git commits and checked-in plans), not through transcript memory.

Evidence: strong

A long task outlives any single context window. Anthropic describes two failures it hit when an agent worked across many sessions: the agent tried to finish everything in one go and ran out of context half way, or a later session saw partial progress and declared the job done (see agent failure modes that grow with one long context). Compaction alone fixed neither, because a summarised transcript is a poor record of what is actually finished.

The remedy that recurs across providers is to put the state of the work into the environment. In Anthropic's pattern an initializer agent writes an init.sh, a claude-progress.txt log and a first git commit, plus a JSON (JavaScript Object Notation) feature list of more than 200 items that all start as failing. Later agents may only flip each item's passes field; JSON was chosen because the model is less likely to rewrite it wrongly than Markdown. OpenAI's cookbook takes the same idea further with "ExecPlans": self-contained living design documents, written for a reader with no prior context, carrying milestones and decision logs, which it credits with letting Codex work for more than seven hours from one prompt. OpenAI's harness post keeps these plans checked into the repository, and Huntley's Ralph loop has the agent record learnings in AGENT.md and bugs in fix_plan.md. Meta's multi-turn recipe keeps a running STATE.md holding the objective, decisions, a task graph with status and a file map, and has the agent consult it before each task.

This is continuity by artefact rather than by runtime journal, a different mechanism from journal-then-replay durable execution, and different from long-term agent memory tiers: the artefacts record the state of one piece of work, not facts to recall later. It is the "write" move in write, select, compress, isolate, and the state a context reset depends on.

The evidence is a strong convergence of lab and practitioner reports rather than a controlled study, so the effect size is unknown.

In practice, an agent's durable record of a task (what was observed, what was tried, what is still open) should live in files or a store it re-reads, not in its own context. Each new session can then reorient through a start-of-session smoke test, take one unit of work per session, and leave the raw event history in a session log outside the context window.

Sources

  • Anthropic Engineering (Justin Young), 26 Nov 2025 — https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
  • OpenAI Cookbook, 7 Oct 2025 — https://raw.githubusercontent.com/openai/openai-cookbook/main/articles/codex_exec_plans.md
  • OpenAI (Ryan Lopopolo), Feb 2026, repository knowledge as system of record — https://openai.com/index/harness-engineering/
  • Geoffrey Huntley, undated — https://ghuntley.com/ralph/
  • Meta Model API cookbook, Agent Patterns, undated (retrieved 3 Oct 2026) — https://dev.meta.ai/docs/cookbook/multi-turn-context-management