Phased Explore-Write-Early Pattern

For a long-running autonomous exploration task, mandate a checkpointed partial write well before completion and periodic updates after — so a useful artifact exists even if the run is cut off early.

An agent given a long, open-ended exploration task (map a codebase and write a development guide; research a topic across dozens of web searches and produce a reference doc) has an obvious failure mode: it accumulates hours of useful findings and never writes them down before running out of budget, hitting a timeout, or being interrupted — total loss of otherwise-good work. The phased explore-write-early pattern removes that risk by making the write step mandatory and early, not a final step the agent might not reach:

  1. Hard early-write deadline. Require the first version of the output file to be written well before the exploration is expected to finish — even if it's incomplete, with explicit placeholder markers (e.g. [TODO]) for sections not yet researched. A rule stated as a hard iteration ceiling ("must call write before iteration 20, no exceptions") is stronger than general advice to "write things down as you go," because it forces the first checkpoint to actually happen rather than being deferred indefinitely by whatever feels more urgent in the moment.
  2. Periodic mandatory updates, not one deferred final write. After the first checkpoint, require the file to be updated at a fixed cadence (e.g. at least every N iterations) as new findings accumulate, rather than holding everything in working memory for a single write at the end.
  3. Priority-ordered sections. When the exploration covers several sub-topics, fix the order they get written in by importance, so a run that terminates partway through still has the highest-value sections filled in rather than an arbitrary subset determined by exploration order.
  4. Citation discipline throughout. Require every claim in the output to carry its evidence inline — a file:line reference for source-code exploration, a source URL for web research — so the output stays independently verifiable rather than becoming an unsourced summary that later readers (including the agent's own downstream consumers) can't check against the original evidence.

This is the crash-resilience analog of the progressive disclosure that Agent Debugger trajectory distillation applies to a reader's consumption of a large trace corpus — applied here to the writer's side instead: rather than structuring an existing large document for cheap partial reads, it structures the production of a long-running task's output so a partial, still-useful document exists at every point in the run, not only at a final write that might never happen. Use it for any agent dispatched to build a durable artifact (a report, a reference doc, a skill file) over a run long enough that early termination is a real possibility — which includes most subagent research or documentation tasks with an open-ended time budget.

Sources

  • Agentic Harness Engineering (Lin, Liu, Pan, et al.), App. B.3