Overnight fanout pacing

On 2026-07-09 Justin asked for an overnight token-max run on CreatorTrack. Two large Workflow fanouts ran concurrently (12-zone audit with Fable finders/fixers plus Opus 3-vote verify, and a CRM research/design/build). Together they exhausted the 5-hour session window in about 75 minutes (~9M tokens in the audit alone), killing 20+ agents mid-flight at 12:45am. Work stalled until the 3:50am reset and finished the next morning.

Why: an overnight burn should span the night, not front-load and die; and daytime capacity must be preserved for Justin's other live sessions.

How to apply: 1. Stagger big workflow launches (or run one at a time) instead of two concurrent token-max fanouts in the same window. 2. Workflow resume works: relaunch with the same scriptPath plus resumeFromRunId; cached find/verify agents replay instantly and only failed legs re-run. Journal at the workflow transcript dir records every agent result. 3. When a loop-until-dry audit reaches diminishing returns (round 3), stop it and synthesize the handoff with one cheap Sonnet agent instead of paying another full verify/fix cycle. 4. Fence concurrent sessions by ownership (files/apps) inside the workflow RULES string; it prevented overlap with the links, ideas-planner, analytics, and perf sessions.

Related: reference_fanout_orchestrator.md, reference_parallel_agent_orchestration.md.

[Claude Code]

URL: https://mkdocs.justinsforge.com/memory/general/feedback_overnight_fanout_pacing/