Skip to content

PATTERN Cited by 1 source

Sequential isolated agent stages with report handoff

Intent

Split a dependent investigation into isolated agent stages and hand off a persistent report after each stage. Separation prevents an early agent’s unsupported conclusion from silently becoming the later agent’s premise, while the report preserves useful evidence across the pipeline.

Structure

reproduce agent
  → report.md: environment, observed behavior, reproduction status
  → diagnose agent
  → report.md: suspected mechanism, evidence, alternatives
  → verify agent
  → report.md: expected-vs-bug determination, relevant tests/docs
  → fix agent
  → patch + regression test + final report

Astro uses this sequence specifically to counter an LLM’s tendency to force a solution even when a reported bug is not real. (Source: sources/2026-08-04-cloudflare-astro-issue-triage)

Design rules

  • Give each stage one decision responsibility and a defined stop outcome.
  • Pass forward durable evidence, commands, artifacts, and uncertainty—not just a natural-language conclusion.
  • Let a later stage reject or amend an earlier conclusion.
  • Keep write authority minimal until evidence has passed verification.
  • Version or retain reports when downstream stages may rerun.

Trade-offs

Isolation improves bias containment and auditability but adds latency and context-management cost. It is inappropriate for independent tasks that can be safely fanned out in parallel; it is a sequential pattern because each stage consumes the previous stage’s evidence.

Seen in

Last updated · 622 distilled / 1,953 read