Agent Swarm

Dependencies

This skill requires Python 3.8+ and standard library only. No external packages needed.

To install this skill's dependencies:

pip-compile ./requirements.in
pip install -r ./requirements.txt

See ../../requirements.txt for the dependency lockfile (currently empty — standard library only).

Agent Swarm

Parallel or pipelined execution across multiple agents and worktrees. The orchestrator partitions work, dispatches to agents, and verifies/merges the results.

When to Use

• Large features that can be split into independent work packages
• Bulk operations (tests, docs, migrations, RLM distillation) that benefit from parallelism
• Multi-concern work where specialists handle different aspects simultaneously

Process Flow

1. Plan & Partition -- Break work into independent tasks. Define boundaries clearly.
2. Route -- Decide execution mode:
   • Sequential Pipeline -- Tasks depend on each other (A -> B -> C)
   • Parallel Swarm -- Tasks are independent (A | B | C) 2.5. Interactively Determine CLI and Model (ask once during bootstrap): Before dispatching the swarm workers, you must ask the user:
   • "Which LLM CLI engine would you like to run the swarm workers through?" (Options: agy, claude, copilot, gemini, llama).
   • "Which specific model should be used?" (Options/defaults per engine, e.g., Gemini 3.5 Flash (Low) or gemini-3.5-flash for agy).
   • Construct the swarm_run.py invocation with --engine and --model matching their choices, appending < /dev/null to prevent TTY input halts (SIGTTIN).
3. Dispatch -- Create a worktree per task. Assign each to an agent:
   • CLI agent (Claude, Gemini, Copilot, Antigravity) using the selected setup
   • Deterministic script
   • Human
4. Execute -- Each agent works in isolation. No cross-worktree communication.
5. Verify & Merge (Trust But Verify & TDD) -- Orchestrator checks each worktree's output against acceptance criteria. No blind trust is allowed.
   • TDD Enforcement: Prioritize running unit and integration tests to ensure no regressions were introduced.
   • Delta Inspection: Check modified files directly for stubs, stales, or placeholders.
   • Verify Quality: If verification fails, generate a correction packet, reject, and re-dispatch.
   • Pass -> Merge into main branch
6. Seal -- Bundle all merged artifacts
7. Retrospective -- Did the partition strategy work? Was parallelism effective?

Worker Selection

Each worktree can be assigned to a different worker type based on task complexity:

Cost Optimization Strategy: For bulk summarization or distillation jobs, use --engine llama (local Gemma 4) if you have local Metal/CUDA acceleration set up. It is the only truly zero-cost path. Cloud CLIs like --engine copilot (gpt-5-mini) or --engine agy (gemini-3.5-flash) are low-cost but paid (consuming AI Credits or per-token billing). Use --workers 2 for cloud CLIs (rate-limit safe) and --workers 1 for local llama to avoid context swapping on 16GB Macs.

Implementation: ./../scripts/swarm_run.py

The ./../scripts/swarm_run.py script is the universal engine for executing this pattern. It is driven by Job Files (.md with YAML frontmatter).

Key Features

• Resume Support -- Automatically saves state to .swarm_state_<job>.json. Use --resume to skip already processed items.
• Intelligent Retry -- Exponential backoff for rate limits.
• Verification Skip -- Use check_cmd in the job file to short-circuit work if a file is already processed (e.g. exists in cache).
• Dry Run -- Test your file discovery and template substitution without cost.
• Engine Flag -- --engine [claude|gemini|copilot|agy] switches CLI backends at runtime.

Usage

# Zero-cost Copilot batch (2 workers recommended to avoid rate limits)
source ~/.zshrc   # NOTE: use source ~/.zshrc, NOT 'export COPILOT_GITHUB_TOKEN=$(gh auth token)'
                  # gh auth token generates a PAT without Copilot scope -> auth failures
python ./scripts/swarm_run.py \
    --engine copilot \
    --job ./resources/jobs/my_job.job.md \
    --files-from checklist.md \
    --resume --workers 2

# Gemini (free, higher parallelism)
python ./scripts/swarm_run.py \
    --engine gemini \
    --job ./resources/jobs/my_job.job.md \
    --files-from checklist.md \
    --resume --workers 5

# Claude (paid, highest quality)
python ./scripts/swarm_run.py \
    --job ./resources/jobs/my_job.job.md \
    [--dir some/dir] [--resume] [--dry-run]

Job File Schema

---
model: haiku        # haiku -> auto-upgraded to gpt-5-mini (copilot) or gemini-3-pro-preview (gemini)
workers: 2          # keep to 2 for Copilot, up to 5-10 for Gemini/Claude
timeout: 120        # seconds per worker
ext: [".md"]        # filters for --dir
# Shell template. {file} is shell-quoted automatically (handles apostrophes safely)
post_cmd: "python ./scripts/my_post_cmd.py --file {file} --summary {output}"
# Optional command to check if work is already done (exit 0 => skip)
check_cmd: "python ./scripts/check_cache.py --file {file}"
vars:
  profile: project
---
Prompt for the agent goes here.

IMPORTANT for Copilot engine: The copilot CLI ignores stdin when -p is used.
Instead, the instruction is prepended to the file content automatically by ./scripts/swarm_run.py.
Do NOT use tool calls or filesystem access - rely only on the content provided via stdin.

Known Engine Quirks

Copilot CLI

• No -p flag -- Copilot ignores stdin when -p is present. ./scripts/swarm_run.py automatically prepends the prompt to the file content instead.
• Auth token scope -- Use source ~/.zshrc to load your token. gh auth token returns a PAT without Copilot permissions, causing auth failures under concurrency.
• Rate limits -- Use --workers 2 maximum. Higher concurrency trips GitHub's anti-abuse systems and surfaces as authentication errors.
• Concurrent writes -- If using a shared JSON post-cmd output (e.g. cache), ensure the writer script uses fcntl.flock for atomic writes. See inject_summary.py.

Gemini CLI

• Accepts -p "prompt" flag normally
• Supports higher concurrency (5-10 workers)
• Model auto-upgrade: haiku -> gemini-3-pro-preview

Checkpoint Reconciliation

If a batch run is interrupted partway through and the output store (e.g. cache JSON) is partially corrupted, reconcile the checkpoint before resuming:

# Remove phantom "done" entries that aren't actually in the output store
completed = [f for f in st['completed'] if f in actual_output_keys]
st['failed'] = {}

Then rerun with --resume.

Constraints

• Each worker execution must be independent
• Post-commands must be idempotent if using resume
• Orchestrator owns the overall job state
• {file} in post_cmd is shell-quoted automatically -- filenames with apostrophes are safe
• Asynchronous Benchmark Metric Capture: Orchestrators MUST capture and log total_tokens and duration_ms from worker agents to a centralized timing.json log immediately as subtasks complete, rather than waiting for the entire swarm batch to finish.

Diagram

See: ./assets/resources/agent_swarm.mmd