docs(agents): one agent per plan task, not one agent per plan

[andybons]

Context-economy rule from today's measurements: monolithic plan-executing agents ran ~700k-800k subagent tokens each (repeated compaction, full-history drag), while fresh per-round reviewers stayed at ~120k-300k and stayed sharp. Plans are zero-context-engineer documents, so per-task dispatch loses no fidelity.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WG6N8YSSwXMma1DKTMwJdf

[claude]

Claude finished @andybons's task in 33s —— View job

---

Code Review

Docs-only change: adds a "One agent per plan task" guidance section to AGENTS.md.

• Read PR diff against origin/main
• Check consistency with surrounding sections and AGENTS.md writing style
• Verify no factual/numeric contradictions
• Post review

---

Verdict: LGTM. No blocking findings.

This is a clean, well-placed documentation addition. It sits logically between "Subagent model strategy" and "Never end a turn to wait on an external event", extending the subagent-dispatch guidance the reader just absorbed.

Strengths

• Follows the repo's ASD-STE100 style: short active-voice sentences, one term reused ("plan task", "dispatch", "agent"), a concrete measured claim (~700k-800k vs ~120k-300k) rather than a vague assertion.
• The measurement date (2026-08-12) is absolute, per the memory/dating conventions.
• The four bullets are actionable and non-overlapping: fresh-per-task, reviewers-per-round, the explicit exception (heavy shared state a plan file can't carry), and file:line pointers.

Minor observations (non-blocking, no change requested)

• The numbers (~700k-800k, ~120k-300k, 2026-08-12) restate the same measurement referenced in the PR body and are self-consistent. Worth keeping in sync if the "Subagent model strategy" section above ever cites the same figures, so the two never drift.
• The exception bullet ("Keep one agent across tasks only when the tasks share heavy state...") is the load-bearing nuance — good that it's explicit, since a strict "always fresh" rule would misfire on live-debugging tasks.

No correctness, concurrency, or protocol surface is touched. Nothing to fix.
· andybons/agents-task-per-agent