Give your coding agent a harness.
Portable skill pack for serious multi-step and multi-agent work — Conductor plans, native memory, hard budgets, external ship gates.

   

              

         

---

agentloop is a portable, high-discipline agent harness — skills, commands, agents, rules, and file-backed state — not a chatbot persona pack and not a commercial visual orchestrator (yet).

It packages prompt → context → harness → loop with:

	Conductor /plan — approved before expensive Execute
	Native local-first memory (.agentloop/memory/ + scripts/memory.sh)
	Doctor diagnostics, handoffs, continuous context surgery, hard budgets
	Competence-bounded council, capped swarm, craft pack that prefers quality

Live site: https://agentloop-ten.vercel.app · Doctrine: docs/doctrine.md · Architecture: docs/architecture.md

┌──────────────────────────────────────────────────────────────┐
│  OVERMIND / CONDUCTOR                                        │
│    ├─ /plan     scaffold · estimates · parallel map · approve│
│    ├─ WARROOM   independent lenses → synthesis (dissent kept)│
│    ├─ PEV       plan → execute → verify (external gates)     │
│    ├─ MEMORY    local store · upsert · supersede · retrieve  │
│    ├─ SWARM     surgical only · hard caps                    │
│    └─ SHIP GATE tests / lint / types / stop conditions       │
└──────────────────────────────────────────────────────────────┘

What shipped

	/plan — Conductor Plan Mode Adaptive scaffold, consumption estimates (turns / context / cost band), parallelization map, agent & model picks, risk register, then a hard approval gate before Execute. Artifacts land in .agentloop/plan.md + stop.yaml.
	Native persistent memory Local-first store at .agentloop/memory/. CLI: scripts/memory.sh — upsert, supersede, contradict, digest, session-brief. Cloud sync field reserved; no external memory service required.
	/doctor + scripts/doctor.sh Surfaces missing setup, hooks, toolchain gaps, and memory-store readiness before you burn a long run.
	Structured handoffs .agentloop/handoff.md carries goal, plan pointer, verify command, progress, risks, next action, and memory_keys — so another session continues cold.
	Continuous context + hard budgets context-surgeon policy (not an occasional skill). Compact on triggers. Honor max_turns, no-progress, and budget aborts in .agentloop/stop.yaml.
	External ship gate Done means tests / lint / types agree — not vibes. Surgical swarm only when Conductor says so.

Install

Surface	Best path	Guide
Cursor	Local plugin or install.sh	adapters/cursor.md
Claude Code	claude --plugin-dir …	adapters/claude-code.md
Codex	install.sh + AGENTS.md	adapters/openai-codex.md
Copilot	Merge into instructions	adapters/copilot.md
Windsurf	Rules + install.sh	adapters/windsurf.md
Agents SDK	Map roles → tools; verify in code	adapters/openai-agents-sdk.md

Bootstrap

git clone https://github.com/edenbuilds/agentloop.git
./agentloop/scripts/install.sh /path/to/your/project

# Optional: also copy commands/, hooks/, and helper scripts
./agentloop/scripts/install.sh --full /path/to/your/project

Default install copies skills + Cursor rules, seeds Conductor-ready .agentloop/ templates (including memory/), and writes AGENTS.md / CLAUDE.md when missing.

--full also copies commands/, .cursor/commands/, hooks/, and helpers (memory.sh, doctor.sh, scaffold-agentloop.sh, …). IDE plugin remains preferred for agents.

After install

1. Run /doctor or scripts/doctor.sh
2. Multi-step work: /plan → approve → /agentloop or /pev
3. Init memory: scripts/memory.sh init · cold starts use session-brief
4. Finish at an external ship gate; keep .agentloop/handoff.md current

Commands

	Command	What it does
	/plan	Conductor Plan — scaffold, budgets, parallel map; waits for approval
	/agentloop	Full harness: plan → loop → verify → ship
	/doctor	Diagnose setup, hooks, toolchain, memory store
	/memory	Native memory — upsert / search / supersede
	handoff	Structured .agentloop/handoff.md for cold continues
	/compact	Context-surgeon compaction + handoff refresh
	/warroom	Council; dissent preserved
	/swarm	Capped parallel workers (surgical only)
	/pev	Plan → Execute → Verify
	/ship	Ship-gate
	/polish /humanise /audit	Craft pack
	/unknowns /quiz	Discovery + merge comprehension

33 skills · 14 commands in this pack.

Memory CLI

scripts/memory.sh init
scripts/memory.sh session-brief
scripts/memory.sh upsert --key my.decision --kind decision --summary "..."
scripts/memory.sh supersede old.key --key new.key --kind decision --summary "..."
scripts/memory.sh contradict a.key --with b.key --note "why"
scripts/memory.sh digest

Doctrine (unchanged)

1. The model is not the product. The harness is.
2. Code owns the loop. The LLM reasons inside nodes.
3. Verify outside the model. Tests, linters, typecheckers, diffs.
4. Delegate surgically. Swarm only for independent, sizeable work.
5. Council before consequential calls. Independent lenses; dissent preserved.
6. Stop conditions beat vibes. Max turns, no-progress, done-when-tests-pass.
7. Finish the asked task. Don't pad, don't quietly widen.

Full write-up: docs/doctrine.md

Agents (lenses, not cosplay)

Agent	Job
overmind	Conductor — plans, budgets, mode, stops
context-surgeon	Continuous compaction + handoffs
first-principles / ship-density / scale-product	Council lenses
verifier	Adversarial external checks
swarm-captain	Spawn/steer/kill with caps
harness-smith	Tools, permissions, hooks, budgets
craft-critic / launch-auditor	Craft + launch readiness

Landing page

Static Vite site in site/:

cd site && npm i && npm run dev

Production: https://agentloop-ten.vercel.app

Repo

https://github.com/edenbuilds/agentloop

Third-party

See THIRD_PARTY.md — includes Finding-Unknowns skills (MIT, © Neeeophytee; technique credit Thariq Shihipar).

License

MIT © edenbuilds