Manager agent + workers in git worktrees: orchestration that survived overnight runs

Fragrant_Yoghurt1135 · reddit · 2026-09-05

A practitioner shares the multi-agent orchestration architecture that survived long unattended overnight runs:

Shape: one long-lived manager agent that plans, allocates, reviews and reconciles but never writes code; workers are headless agent processes each pinned to its own git worktree (one writer per worktree). Every unit of work is a brief written to disk before dispatch; workers write findings to unique report files, with the returned message only a convenience copy.

Lessons only visible unattended:

Silence is not "no findings": one night 6 of 10 workers returned nothing, yet all work existed on disk, one having completed an entire phase gate.
Exit codes lie both ways; only the report file's status line is trusted. Log-grepping ledgers rot — read the artifact.
Concurrency cap: at most 2 workers at once, beyond which they contend for test runners/dev ports; never kill processes by name pattern.
Prompt budgets are advisory; an external watchdog polls session files, counts turns and terminates past a cap.
Adversarial verification: the writer never verifies its own change; verification and staged-diff scope review go to agents with no knowledge of the change — more effective than any prompt engineering.

Open problems: manager context as multi-day bottleneck, unit-of-work sizing, drift detection, and automating the rule-improvement loop.

More from coding & agent