What
A parallel agent fleet
N autonomous Claude sessions, one per independent group of work, each isolated in its own git worktree and branch, coordinated by a master.
Run a fleet of autonomous Claude Code agents in parallel — one per independent chunk of work, each a full claude session in its own git worktree, coordinated by a changeable master through a command centre that lives on disk. The operation survives context limits and shift changes: any session can read the command centre and take command.
It's the heavyweight, separate-session way to parallelise in Claude Code. Instead of one chat grinding through a backlog, you spawn N full claude processes at once — each in its own terminal window/tmux pane, its own context window, its own git worktree — and a master session coordinates them through an on-disk command centre. Because the command centre is just files on disk (outside any repo), the whole operation outlives any single chat: a fresh session can read it and take command.
agent-army isn't a rival to subagents — it sits one level up. Each fleet agent runs its own in-session subagents to keep its context lean.
What it is, why it exists, when (and when not) to reach for it, who runs it, where its state lives, and how a run actually goes.
N autonomous Claude sessions, one per independent group of work, each isolated in its own git worktree and branch, coordinated by a master.
It buys three things nothing else does: N independent context windows, hours of true parallel building, and an operation that outlives any one chat. It trades ~N× the spend for that.
Use it when work is already split into 4–9 substantial, independent streams, there’s a real deadline, and a human will babysit. Not for small or coupled work — the Intake Gate routes those away.
A master coordinates (spawns, monitors, unblocks, reviews, assembles — writes no feature code). The agents are semi-autonomous workers, one per seam. The master is swappable mid-run.
State lives in ~/.claude/agent-army/<operation>/ — outside any repo, referenced by absolute path. Agents work in their own worktrees and PR into one integration branch.
Run the Intake Gate, cut an integration branch, build the command centre, launch + verify each agent, monitor with two-signal health, then review and assemble one final PR.
Six phases, master-driven. It’s semi-autonomous — the agents build on their own, but you check in to clear blocked decisions.
Confirm the independent groups of work and each group’s spec (an OpenSpec change or a markdown brief). One seam = one agent = one branch. Target 4–9.
Branch off the agreed base; every agent PRs into it (never into main). Choose models per agent — sonnet is the cost/quality sweet spot when specs are written.
Scaffold ~/.claude/agent-army/<operation>/ with COMMAND.md, manifest.json, and per-agent brief/status files. Give the operation a memorable codename.
Launch one agent per seam. Verify each is actually live within ~15s — never trust the launch exit code. Each agent self-creates its worktree as step one.
scripts/launch-agent.sh <operation> <agent>Read the objective health board; locate windows; clear BLOCKED questions; relaunch by reason; stop anything idle-burning.
scripts/fleet-status.sh <operation>scripts/where.sh <operation>When agents are PR-OPEN, review each PR two-stage (spec-compliance, then code-quality) with in-session reviewer subagents, merge into the integration branch, run CI, open one final PR, then stop the fleet and archive the command centre.
Agents stall silently at usage limits, commits lag reports, and launches lie about success. So health is two signals together: what the agent reports, and an objective probe.
--resume — context is fine, just frozen at a usage limit.stop-agent.sh it. That’s real money.A fleet costs roughly N full sessions. The skill opens with an Intake Gate that does this maths and routes cheap work to a cheaper tool. Use this rule first.
| Shape of the work | Use |
|---|---|
| 4–9 big independent seams, need judgement + review, span hours/days, human babysits | agent-army |
| Many small, mechanical, isolated tasks (rename, dep bump, header sweep) | Workflow tool |
| A handful of independent tasks you’ll finish in one sitting | In-session subagents |
| Tightly coupled work (same files / ordered dependencies) | Sequential |
Add the marketplace once, then install the skill. Trigger it by asking for a fleet — it always runs the Intake Gate first, so it’ll talk you out of a fleet you don’t need.
/plugin marketplace add amjad1233/claude-skills/plugin install agent-army@amjad1233