Agence

l'Agence — Agentic Engineering Co-Environments

Author: Stephane Korning · 2026 · MIT + Commons Clause Version: v1.0.0 · May 2026

The governance layer for AI coding agents.

Every agent action classified, gated, and cryptographically logged — regardless of which LLM or tool runs it.

---

The Problem

Your AI coding agents can write code. They can commit, push, delete, refactor.

But who's watching them?

Claude Code has no audit trail. Aider trusts the user. Codex sandboxes everything and hopes for the best. LangChain gives you building blocks but no guardrails.

Agence exists because advisory guardrails aren't guardrails at all. It's the layer that sits between any AI agent and your filesystem and says *"not without approval."*

---

What Agence Does

Agence is an agent-agnostic governance stack for software engineering. It doesn't replace your coding agent — it governs, orchestrates, and audits all of them from a single control plane.

Command Gating — Every shell command is classified before execution:

Unknown commands default to T2. Not T0. Fail-closed. The guard runs as a separate process — agents cannot bypass their own policy.

Cryptographic Audit — Every agent decision is logged to a Merkle-chained, append-only ledger. Each entry links to the previous via SHA-256. Tamper with one entry and the chain breaks. Verify with: agence ^ledger verify.

Multi-Agent Orchestration — 18 agents across 4 types (persona, tool, loop, ensemble). Route with @agent syntax. Override models with dot-notation: @ralph.gpt4o. Dispatch to Aider, Claude Code, Copilot, or your own tools — all governed by the same policy.

Peer Consensus — Route any question to 3 independent LLMs and get weighted consensus. Your architecture review shouldn't depend on one model's blind spots.

Session Persistence — Save, resume, and hand off sessions between agents. Full context survives restarts. Automatic tmux capture of stdout/stdin/stderr — no 16KB buffer limits.

Git-Native — No database. No server. State lives in git worktrees and flat files. Knowledge is sharded, gated, and selectively routed — you decide what gets shared.

---

By the Numbers

---

Who This Is For

• Teams using multiple AI coding agents who need one policy governing all of them
• Enterprises requiring audit trails for AI-generated code changes
• Security-conscious developers who want fail-closed gating, not fail-open trust
• Anyone who's had an AI agent break something and wished there was a layer between the agent and rm -rf

Who This Is NOT For

• If you want an AI pair programmer → use Aider
• If you want IDE autocomplete → use Continue or Copilot
• If you want to build any kind of agent → use LangChain/LangGraph
• If you want cloud-hosted async tasks → use OpenAI Codex

Agence governs all of the above.

---

GitHub Action — Agence Guard

Gate AI agent commands in any CI workflow — one step, zero infrastructure:

- name: Gate AI command

id: guard uses: l-agence/agence@v1 with: command: 'git push origin main' # command proposed by your AI agent agent: ci fail_on_block: 'true' # fail if T3-blocked

• run: echo "Tier ${{ steps.guard.outputs.tier }} — ${{ steps.guard.outputs.reason }}"

See docs/marketplace/description.md for full documentation.

This PR starts with the GitHub Action path first for fast per-repo adoption; the GitHub App listing and webhook server can follow in a later phase or separate PR for org-level rollout across many repositories.

For enterprise adoption, the next layer after the GitHub App is shard onboarding so orgs can separate teams, policies, and knowledge boundaries cleanly instead of forcing one shared shard model.

---

Install

As a git submodule (recommended)

git submodule add https://github.com/l-agence/agence .agence

git submodule update --init --recursive bash .agence/bin/agence ^init export PATH="$PWD/.agence/bin:$PATH"

Or: clone directly

git clone https://github.com/l-agence/agence .agence

cd .agence && bun install ./bin/agence ^init export PATH="$PWD/.agence/bin:$PATH"

Prerequisites

Windows: Use WSL (Ubuntu recommended).

---

Quick Start

# Chat with an agent

agence "How should I structure this feature?"

Route to a specific agent

agence @sonya "Review this auth module"

Launch an agent shell

agence !ralph # Persona: autonomous iteration agence !claude # Tool: Claude Code CLI agence !aider # Tool: aider (code patches)

Save session (resume later or hand off to another agent)

agence ^save "OAuth2: done token validation, next: refresh flow" agence ^resume agence ^handoff @sonya

Audit trail

agence ^ledger verify # Verify Merkle chain integrity agence ^audit trail # View full decision history

Peer consensus (3 independent LLMs)

agence @peers "Should we use Redis or Postgres for session storage?"

See all commands

agence --help

---

Architecture

YOUR REPO/

└── .agence/ ← lives here (submodule or clone) ├── bin/ # CLI: agence, aibash, ibash, aido, agentd ├── codex/ # Governance: AIPOLICY.yaml, Laws, Principles, agents/ ├── nexus/ # Local state: .ailedger, sessions, faults (gitignored) ├── knowledge/ # Team knowledge: docs, lessons, plans (committed) │ └── private/ # Private knowledge (gitignored, never shared) ├── organic/ # Swarm coordination: tasks, jobs, workflows └── lib/ # Core: guard.ts, signal.ts, skill.ts, memory.ts, peers.ts

COGNOS — Four pillars:

Runtime: Bun + bash. No Python. No pip. No npm install of untrusted packages in the critical path.

MCP: Agence exposes itself as an MCP server (10 tools, 3 resources) so any MCP-compatible client can use agence's governance layer. Agence also acts as an MCP client — consuming tools from external MCP servers. See MCP.md for integration guide.

---

Command Reference

---

Agent Roster

Override models with dot-notation: @ralph.gpt4o, @sonya.opus, @ralph.aider

---

Governance

Agence uses a 5-tier command policy. The guard runs as a separate process — agents cannot bypass their own policy.

Unknown commands default to T2. Fail-closed. 120+ rules across git, GitHub CLI, AWS, Terraform, and shell.

All decisions logged to nexus/.ailedger — append-only, Merkle-chained, HMAC-signed.

See SECURITY.md for full security architecture, red-team findings, and disclosure timeline.

---

Swarm (agentd)

agentd start ralph claude aider   # Launch 3 agents in tmux

agentd tangent create fix-auth # Isolated worktree + container agentd inject fix-auth "run tests" # Send command via socat socket agentd status # View all agents + tangents

Each tangent gets: isolated git worktree, optional Docker container (--cap-drop ALL, --read-only, --no-new-privileges), socat socket for IPC, tmux pane for observability.

---

Tests

bun test                          # Full suite

751 tests, 1,768 assertions, 0 failures across 21 files:

| mcp-client.test.ts