Make intent verifiable.
A clear contract. A small change. Evidence you can inspect. MADD helps humans and agents agree on what should happen, then check what actually shipped.
.madd/contract.d/
- 01 / requirement
- REQ-F-001
- 02 / check
- TEST-001
- 03 / binding
- planned
- 04 / status
- pending
A written promise. Evidence comes next.
A change with a receipt
Think boldly. Ship something you can explain.
MADD means Multi-Agent Driven Development. Start with one developer, one independent reviewer and your existing CI. Add agent roles when they help the work.
Try one small change · Read the contract
intent → one observable outcome
contract → requirement + acceptance check
fraction → one reviewable change
checks → actual execution results
review → an independent challenge
retro → what shipped, and what did not
Start with the change, not the prompt
A long prompt is a poor handover. Put one observable outcome, its limits and its acceptance check in versioned files. The developer implements it, CI records what ran, and an independent reviewer checks whether the result matches the promise.
One change, six handoffs
Architect clarifies intent. Maker implements. CI executes checks. Breaker challenges the change independently. Witness records observed behavior. Conductor keeps the scope and handoffs clear. These are responsibilities, not a requirement to run six AI processes.
Small fractions, explicit limits
A fraction is a small delivery scope with named requirements and checks. A green test does not prove every property of a system. A review is not a guarantee of zero defects. MADD makes those boundaries visible so the next person can continue safely.
Start with what is available
The published 0.1.3 CLI installs optional agent workflows. Its doctor command checks installation health.
npx @madd-sh/[email protected] init
The 0.2.0-rc.1 candidate adds contract-only initialization, structural validation and evidence verification. Try it from npm with npx @madd-sh/madd@next. Read the candidate workflow.
Keep your tools
Use the repository, language, CI and review tools your team already understands. Adopt the contract first. Explore integrations, inspect evidence, or read without a browser.
Compare observed results
Use your own baseline, scope and checks to judge whether the method helps. Our evidence policy explains what a reproducible comparison needs.