Harness engineering for agent swarms that ship.
Council. Loop. Verify. Done.
GODMODE is a portable agent plugin — not a chatbot persona pack. It packages the 2026 stack that actually moves product: prompt → context → harness → loop, with a competence-bounded council and a capped subagent swarm.
Works across Cursor, Claude Code, OpenAI Codex, GitHub Copilot, Windsurf, and OpenAI Agents SDK-style orchestrations.
┌─────────────────────────────────────────────────────────┐
│ OVERMIND (orchestrator) │
│ ├─ WARROOM council (independent lenses → synthesize) │
│ ├─ SWARM (parallel scoped workers, hard spawn caps) │
│ ├─ PEV loop (plan → execute → verify externally) │
│ └─ SHIP GATE (tests / lint / diff budget / stop) │
└─────────────────────────────────────────────────────────┘
- Clone this repo (or copy the
godmode/folder). - Add as a local plugin from Customize → Plugins, or symlink into your Cursor plugins path.
- Confirm agents/skills appear; try
/warroomor/godmode.
claude --plugin-dir /path/to/godmode
# or install from a marketplace entry pointing at this directory# skills auto-discover from skills/*/SKILL.md
# also loads via .agents/plugins/marketplace.json
cp -R godmode/.agents your-project/.agents # optional marketplace pointer
cp godmode/AGENTS.md your-project/AGENTS.md # or merge sections./scripts/install.sh /path/to/your/project| Command | What it does |
|---|---|
/godmode |
Full harness mode: scope, loop, verify, ship |
/warroom |
Convene the council; preserve dissent |
/swarm |
Spawn capped parallel subagents |
/pev |
Plan → Execute → Verify until external gates pass |
/ship |
Run ship-gate: tests, lint, scope, stop |
- The model is not the product. The harness is.
- Code owns the loop. The LLM reasons inside nodes; it does not invent control flow.
- Verify outside the model. Tests, linters, typecheckers, diffs — not "are you sure?"
- Delegate surgically. Swarm only for independent, sizeable, parallelizable work.
- Council before consequential calls. Independent lenses first; synthesis second; dissent preserved.
- Stop conditions beat vibes. Max turns, no-progress, escalation, done-when-tests-pass.
- Finish the asked task. Don't pad, don't quietly widen, don't over-verify theater.
Full write-up: docs/doctrine.md · Sources: docs/resources.md
These are methodological roles. They are not celebrity impressions and never claim endorsement.
| Agent | Job |
|---|---|
overmind |
Lead orchestrator; owns the loop and stop conditions |
first-principles |
Strip folklore; design the smallest true loop |
ship-density |
Cut scope theater; maximize shipped surface area |
scale-product |
Product clarity, distribution, AGI-ready systems thinking |
verifier |
Adversarial external checks; never grades its own homework alone |
swarm-captain |
Spawn/steer/kill workers with lineage + caps |
context-surgeon |
Compaction, handoffs, KV hygiene |
harness-smith |
Tools, permissions, hooks, sandboxes, budgets |
godmode-core— master orchestrationwarroom-council— independent lenses → synthesisharness-design— six-layer harness buildagent-loop— think/act/observe/verify with hard stopssubagent-swarm— parallel workers without chaospev-cycle— Plan-Execute-Verifycontext-compact— multi-stage context managementopus-prompting— Claude Opus 5 agentic prompting patternsship-gate— external verification before "done"writer-verifier— separated author and checker
This package currently lives under edenbuilds/bloom while the GitHub App token for this agent only has write access to that repository.
To publish as edenbuilds/godmode:
./scripts/extract-standalone.sh
# requires: gh auth with repo-create on edenbuildsMIT © edenbuilds