agent-army

Deploy a two-layer agent hierarchy: a commander model plans the work, then 3-50+ Layer 1 agents each spawn 2+ sub-agents to edit files in parallel. A deployment gate and git diff verification confirm every file was actually modified.

A swarm splits one context window across sub-agents, while an army gives each Layer 1 agent its own full context window. The army approach suits large tasks where each agent needs deep independent context rather than a divided slice.

Avoid it for fewer than 6 independent units of work, single-file changes, heavily sequential dependencies, or tasks needing one coherent authorial voice. Orchestration overhead makes small jobs slower and burns extra tokens.

Pick a power level: Max Power (Opus everywhere), Heavy (Opus L1, Sonnet L2), Balanced (Sonnet), or Economy (Sonnet L1, Haiku L2). Always pass the model parameter explicitly on every agent call to avoid silent inheritance of the expensive session model.

Run git diff --stat and cross-reference it against each agent's completion report. An agent reporting files modified with no matching diff is a phantom completion and should be re-dispatched. Trust the diff, not the report.

Yes. The skill writes an .army-state.md scratchpad after Wave 1 recording diff-confirmed files, open flags, and decisions. A new invocation reads this file and resumes from the last clean wave instead of restarting.