# Coordinate Autonomous Coding Agents

Complex coding work can lose context across many implementation and review steps. Swarm coordinates iterative agents through a shared plan, review gates, tests, and completion checks.

## Install

```bash
npx skillstore add nbashaw/swarm
```

## Metadata

- Status: approved
- Slug: nbashaw-swarm
- Skillstore revision: r2
- Version status: missing
- Tree hash: cc7e9b45971ed70a470f433e7a14ae90a5bbef29696e2c9e6ecf8d74fa8d2198
- Author: nbashaw
- GitHub username: nbashaw
- License: MIT
- Repository: https://github.com/nbashaw/swarm/tree/main/swarm
- Ref: 02be9409c79ca1183f7844009c14d9df684d0cf9
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: confirmation\_required
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: filesystem, external\_commands
- Quality score: 38
- Quality tier: warning
- Public page: https://skillstore.pages.dev/skills/nbashaw-swarm
- Manifest: https://skillstore.pages.dev/api/skills/nbashaw-swarm/manifest

## Capabilities

- Interviews users to define goals, scope, plan location, and verification commands.
- Creates Markdown plans with phases, checkboxes, role changes, and a shared progress log.
- Runs sequential Codex or Claude agent iterations through a Bash script.
- Adds reviewer, simplifier, tester, and gatekeeper stages to the execution plan.
- Supports dry-run, verbose, agent selection, and configurable iteration limits.
- Uses a project-local completion marker to stop the workflow.

## Use Cases

- Refactor a subsystem: Coordinate implementation, simplification, tests, and final verification across several files.
- Migrate a codebase: Track a staged framework or API migration with explicit checks after each phase.
- Expand test coverage: Plan coverage work, run targeted tests, review gaps, and verify completion evidence.

## Prompt Templates

### Plan a coding task

```
Use Swarm to plan [task]. Ask me for the goal, affected files, test commands, and preferred plan location before creating the plan.
```

### Execute an approved plan

```
Use Swarm to execute [plan file]. Start with a dry run, summarize the command, and wait for my approval before launching agents.
```

### Coordinate a migration

```
Create a Swarm plan for migrating [system] from [current state] to [target state]. Include rollback points, compatibility tests, and review gates.
```

### Run a constrained autonomous workflow

```
Plan and execute [goal] within [directories]. Use [test commands], preserve [constraints], stop on unexpected changes, and provide gatekeeper evidence before completion.
```

## Limitations

- Agents run sequentially rather than concurrently.
- The runner requires Bash and an installed Codex or Claude command-line client.
- Autonomous modes reduce permission prompts and should run only in an isolated worktree or container.
- Plan quality and repository instructions directly affect execution quality and safety.

## Best Practices

- Use measurable completion criteria and exact verification commands.
- Review the generated plan and dry-run output before starting autonomous execution.
- Run the workflow in an isolated worktree or container with limited credentials.

## Anti Patterns

- Do not use Swarm for typos, single-line fixes, or tasks requiring frequent decisions.
- Do not launch agents against an unreviewed plan or untrusted repository instructions.
- Do not treat completion markers as proof without checking tests and gatekeeper evidence.

## Security Audit

- Audited at: 2026-08-09T10:04:17.187\+00:00
- Summary: Most static findings are false positives caused by Markdown backticks, shell arithmetic, test commands, and local completion markers. One temporary-file issue is confirmed, and semantic review found a high-risk autonomous execution path that runs agents with reduced or disabled permission controls.

## Stats

- Views: 243
- Downloads: 13
- Favorites: 1
- Popularity score: 0