Modular AI agents.
Built to scale. Quilt to ship.

A Git-native framework for AI agent instructions as small, ordered fragments. Edit structured source, then compile deterministic outputs — .claude/agents, AGENTS.md, Agent Skills — with a lock file that catches drift in CI.

View on GitHub
agentquilt
.agentquilt/
config.yaml
agents/
_shared/
reviewer/
agent.yaml
010-role.md
020-style.md
skills/
.claude/agents/ gen
agentquilt.lock gen
COMPILES FOR
Claude Code Agent Skills + more soon
01

Modular by design

Author agents as small, ordered instruction blocks — plus shared fragments reused across agents.

02

Fewer merge conflicts

Tiny files edited in parallel collide far less than one giant hand-maintained prompt.

03

CLI-first

One tool to init, build, check and scaffold agents & skills.

04

Multi-platform

Claude Code and Agent Skills adapters ship today. Cursor, Copilot, Gemini, and more coming soon.

05

Deterministic & drift-checked

Content-hashed outputs and a lock file. check fails CI on any stale or hand-edited file.

Why AgentQuilt

One big agent file. A whole team editing it. Constant conflicts.

Large agent instruction files are hard to maintain in distributed teams. Two people editing the same prompt hit merge conflicts that are painful to resolve — agent instructions are semantic, not purely syntactic. And the same rules, copied by hand into Claude’s CLAUDE.md, OpenAI’s AGENTS.md, and more, quietly drift apart.

source fragment · _shared/040-review-policy.md

“All PRs require two approvals and a green CI run before merge.”

copied into 4 platform files ↓

CLAUDE.md Claude Code

“Require two approvals before merge.”

partial · 2d ago
AGENTS.md OpenAI Codex

“One approval is fine — move fast.”

contradicts · 3w ago
reviewer.md Claude subagent

“Two approvals, plus a security pass.”

reworded · 5w ago
SKILL.md Agent Skills

“(review policy not mentioned)”

omitted · 2mo ago

How it works

Quilt once.
Run everywhere.

Each agent is a small agent.yaml manifest plus ordered instruction blocks. agentquilt build validates them and compiles deterministic outputs for every configured platform — writing a lock file so CI can catch drift.

See the source model →

source fragments · .agentquilt/agents/reviewer

agent.yaml
010-role.md
020-style.md
_shared/
>_ agentquilt build

compiled outputs · never hand-edited

CLAUDE.md
AGENTS.md
.claude/agents/*.md
.agents/skills/SKILL.md
+ Cursor, Copilot, Gemini soon

The source model

A manifest, some blocks, a generated agent.

Agent = Manifest + Instruction Blocks + Generated Prompt

.agentquilt/agents/reviewer/agent.yaml
description: Code reviewer for pull requests
model: balanced  # frontier | balanced | fast
permissions: workspace  # read-only | workspace | full
x-claude:
  tools: [Read, Grep, Edit]
.agentquilt/agents/reviewer/020-style.md
---
tags: [style, review]
---
## Style
- Prefer composition over inheritance.
- Keep functions under 40 lines.
.agentquilt/config.yaml
version: 1
sourceDir: .agentquilt/agents
defaultModelTier: balanced
modelTiers:
  balanced: { claude: claude-sonnet-4-6 }
targets:
  - output: AGENTS.md  # doc target
  - kind: agent-definitions
    agents: [reviewer]
    platforms: [claude]
.claude/agents/reviewer.md (generated)
---
name: reviewer
description: Code reviewer for pull requests
model: claude-sonnet-4-6
---
# compiled by agentquilt · do not edit
## Style
- Prefer composition over inheritance.

Quickstart

Author → build → check.
In under a minute.

Install once, scaffold in your repo, then commit sources and generated outputs together. Node ≥ 18 · open source · MIT.

quickstart
1 $ npm install -g agentquilt
2 $ agentquilt init --platform claude
3 $ agentquilt agents add reviewer
4 $ agentquilt build  # → targets + lock