Agentic Sprint brings structured, spec-driven development to Claude Code. A self-iterative state machine that reads your specs, orchestrates agents, and loops until the work is done.
A spec-driven, self-iterative state machine in four phases
Create a simple specification file describing what you want to build. Can be one line or highly detailed.
The Project Architect agent analyzes your specs and creates detailed API contracts and implementation guides.
Specialized agents (backend, frontend, CI/CD) implement the feature in parallel, following the specs.
QA and UI tests validate. The state machine loops — refining specs, fixing issues — until convergence (max 5 iterations).
Start with built-in agents, extend with your own. The architecture is open by design.
Analyzes requirements, creates specifications, and coordinates all other agents. The decision-maker.
CoordinationBuilds FastAPI backends with async patterns, PostgreSQL, and proper security practices.
BackendCreates modern React applications with Next.js 16, TypeScript, and TailwindCSS.
FrontendMaintains automated test suites with API and unit tests. Validates against the API contract.
TestingRuns E2E browser tests via Playwright MCP. In manual mode, opens a live browser for you to explore while capturing errors.
E2E TestingConfigures pipelines, manages deployments, and sets up quality gates.
DevOpsNo specialized agent for your stack? The allpurpose-agent adapts to any technology — Go, Rust, Flutter, Ruby, whatever. The architect automatically falls back to it when needed.
The architect creates appropriate specs and prompts it correctly. You just write your requirements.
These agents are starting points, not final products. Fork, extend, contribute. Add agents for your stack, improve existing ones, share what works.
Learn how to create agents →A convergent process that eliminates AI's common failure modes
Each agent receives only what it needs — specs, contracts, relevant code. No wasted tokens on irrelevant history. Context windows stay lean and focused.
Completed work is removed from specifications. Each iteration focuses only on what remains. TODOs decrease, clarity increases.
Working code stays untouched while issues get fixed. The signal-to-noise ratio improves with each pass. Mistakes don't compound.
Think of it like a diffusion process: the picture starts noisy, but with each iteration, noise reduces and clarity emerges. Most sprints converge well before 5 iterations. If not, the system pauses and asks what to do — you stay in control.
Persistent memory that survives across sprints — without bloating your context
The Business Brain (you maintain this)
The architect reads this to make decisions aligned with your product goals, not just technical specs.
The Technical Brain (architect maintains)
Agents read this instead of scanning the entire codebase. Straight to the relevant files.
You maintain project-goals.md — the more detail you provide, the sharper the architect becomes. The architect maintains project-map.md and keeps it lean. This is how agents "remember" your project without consuming your context window.
Everything you need for autonomous, spec-driven development
API contracts are the source of truth. Frontend and backend implement the same spec.
Autonomous iterations with a checkpoint at 5. System pauses to ask — continue, adjust, or intervene.
Implementation agents run in parallel. UI testing and diagnostics run together.
Leverages Playwright and Next.js DevTools MCP for browser automation and error monitoring.
Every agent returns machine-parseable reports. Easy to track progress and issues.
French and English support throughout. Agents use i18n keys, never hardcoded strings.
Explore your app with live error monitoring
A real browser opens pointing to your app. Click around, test forms, explore edge cases yourself.
A diagnostics agent watches for runtime errors in real-time. We ship one for Next.js — write your own for any stack with observability tools.
When you close the browser, both agents finalize and return comprehensive reports of everything captured.
Run /sprint --manual to skip the full workflow and jump straight to manual testing.
All you need to start a sprint
# Sprint 1 Specifications ## Goal Add user authentication with OAuth and email/password ## Scope - User registration and login - OAuth with Google and GitHub - Password reset flow - Protected dashboard route ## Testing - QA: required - UI Testing: required - UI Testing Mode: automated
Clone the repo, copy to your project, and run your first autonomous sprint.
Get Started on GitHub →