Skip to content

Multi-agent autonomous development orchestrator powered by GitHub Copilot CLI

Notifications You must be signed in to change notification settings

markgar/buildteam

Repository files navigation

Multi-Agent Flow

An autonomous development workflow using GitHub Copilot CLI. Six agents collaborate through files in a shared GitHub repo to build projects in .NET/C#, Python, or Node.js:

  • Planner — creates a backlog of stories and expands them into milestones one at a time
  • Builder — fixes bugs, addresses review items, completes one milestone per cycle
  • Commit Watcher — reviews each commit for quality, fixes doc issues directly, files code issues as GitHub Issues
  • Milestone Reviewer — cross-cutting milestone reviews with frequency-filtered findings and stale-item cleanup
  • Tester — runs scoped tests when milestones complete, files bugs
  • Validator — builds the app in a Docker container, runs it, and validates it against the spec

Prerequisites

Install these one time. After each install, close and reopen your terminal so the PATH updates.

Required for all projects:

Tool Install (Windows) Install (macOS) Verify
Git winget install Git.Git brew install git (or Xcode CLT) git --version
GitHub CLI winget install GitHub.cli brew install gh gh auth status
Python 3 winget install Python.Python.3.12 brew install python python3 --version
Docker winget install Docker.DockerDesktop brew install --cask docker docker --version
Copilot CLI (coding agent) See Copilot coding agent on CLI Same copilot --version

Install for your target language:

Language Tool(s) Install (Windows) Install (macOS)
dotnet .NET SDK winget install Microsoft.DotNet.SDK.9 brew install dotnet
python Python 3 (already installed above) (already installed above)
node Node.js winget install OpenJS.NodeJS.LTS brew install node

After installing GitHub CLI, authenticate:

gh auth login

Quick Start

git clone https://github.com/markgar/buildteam
cd buildteam
pip install -e .
cd ..

buildteam go --directory hello-world --model gpt-5.3-codex \
  --description "a C# console app that prints Hello World"

That's it. go will:

  1. Bootstrap — create the project, SPEC.md, git repo, GitHub remote, and agent clones (builder-1/, reviewer-1/, milestone-reviewer/, tester/, validator/)
  2. Plan — generate BACKLOG.md (story queue)
  3. Launch commit watcher — in a new terminal window, reviewing each commit
  4. Launch milestone reviewer — in a new terminal window, cross-cutting reviews when milestones complete
  5. Launch tester — in a new terminal window, testing each completed milestone
  6. Launch validator — in a new terminal window, building containers and validating against the spec
  7. Start building — in your current terminal, completing milestones one at a time

Multiple windows will be running. Watch the Builder work through milestones while the Reviewer, Tester, and Validator react to each commit and milestone completion. Run buildteam status anytime in a builder directory to check progress.

Continuing with New Requirements

Come back later and add new features to the same project:

buildteam go --directory hello-world --model gpt-5.3-codex --spec-file new-features.md

The planner detects what's already built, updates the spec, and creates new milestones for the unimplemented work.

Resuming Where You Left Off

buildteam go --directory hello-world --model gpt-5.3-codex

No spec needed — just re-evaluates the plan and continues building.

Resuming from a Different Location

Point --directory at any existing project directory, even from a test harness run:

buildteam go --directory /path/to/runs/20260213/my-app --model gpt-5.3-codex

go detects the existing repo on GitHub via gh repo view and automatically clones any missing agent directories. You can resume on a fresh machine with nothing but the repo — no pre-existing builder-1/, reviewer-1/, tester/, or validator/ directories needed.

GitHub Org

Create the repo under a GitHub organization instead of your personal account:

buildteam go --directory my-project --model gpt-5.3-codex --org my-org --description "..."

Per-Agent Model Overrides

Use different models for different agents:

buildteam go --directory my-project --model claude-opus-4.6 \
  --builder-model gpt-5.3-codex \
  --planner-model claude-sonnet-4.6 \
  --description "..."

Available overrides: --builder-model, --reviewer-model, --milestone-reviewer-model, --tester-model, --validator-model, --planner-model, --backlog-model. Each defaults to --model when not specified.


Examples

See EXAMPLES.md for ready-to-run commands — Hello World, REST API, and Full Stack projects in .NET, Python, and Node.js.


How It Works

Planner ──milestones/──→ Builder ──git push──→ Reviewer ──GitHub Issues──→ Builder
                           ↑                                              
                           │         Builder ──git push──→ Tester
                           │                                  │
                           │         Builder ──git push──→ Validator
                           │                                  │
                           └───── GitHub Issues ───────────┘
From To Mechanism What it says
Bootstrap Planner SPEC.md Here's the technical decisions for how to build the project
Planner Builder BACKLOG.md, milestones/ Here's the next milestone to build
Builder Reviewer git push I finished a commit or milestone, review it
Builder Tester milestones.log A milestone is complete, test it
Builder Validator milestones.log A milestone is complete, validate it in a container
Reviewer Builder GitHub Issues (finding) I found code-level issues, address these
Milestone Reviewer Builder GitHub Issues (finding) Recurring patterns promoted from notes
Reviewer (self) direct commit I found a doc issue (stale comment, inaccurate README), fixed it myself
Tester Builder GitHub Issues I found test failures, fix these
Validator Builder GitHub Issues The app failed validation in a container, fix these
Validator Builder DEPLOY.md Here's what I learned about deploying this app

Positive Feedback Loops

The system is designed around cumulative knowledge — each milestone makes subsequent milestones more reliable:

  • Deployment knowledge (DEPLOY.md): The validator writes everything it learns about building and running the app (Dockerfile config, env vars, ports, startup sequence, gotchas). The builder reads it to stay compatible. Each milestone's validation inherits all prior knowledge, so deployments get more reliable over time.
  • Review signal filtering (notes → findings): Per-commit reviews file [bug]/[security] issues immediately but file [cleanup]/[robustness] issues as observational notes. The milestone review then evaluates notes for recurring patterns — only issues that appear in 2+ locations get promoted to findings for the builder. This means the builder spends time on systemic problems, not one-off nitpicks.
  • Review themes (REVIEW-THEMES.md): The reviewer maintains a rolling summary of the highest-impact recurring patterns. The builder reads it to avoid repeating the same class of mistake across milestones.
  • Codebase-aware planning: The milestone planner reads the actual codebase before expanding the next story — it matches existing patterns (base classes, naming conventions, DI wiring) rather than planning in a vacuum.
  • Copilot instructions (.github/copilot-instructions.md): The builder updates this style guide as the project structure evolves. Future builder sessions read it, so coding conventions stay consistent as the project grows.

See AGENTS.md for the exact prompts each agent receives and coordination rules.

See USAGE.md for the full command reference, logging, troubleshooting, and more.

See docs/docker.md for running in a Docker container (headless, CI, or server environments).

About

Multi-agent autonomous development orchestrator powered by GitHub Copilot CLI

Resources

Stars

Watchers

Forks

Packages

 
 
 

Contributors