An autonomous AI agent scaffold for GitHub Copilot that turns a written goal into a working project — using GitHub Copilot subagents to plan, execute, and deliver.
You describe what you want built, preferably in user stories. Skillwave handles the how: it decomposes the goal into tasks, creates specialized worker roles on demand, dispatches up to 4 parallel subagents, and loops until everything is done.
- You write a goal in
GOAL.md— requirements, constraints, success criteria. - Invoke
@skillwavein copilot. - Skillwave takes over — it reads your goal, creates a task board, spins up roles (e.g. backend engineer, designer), and delegates work to subagents.
- Work happens autonomously — tasks execute in parallel, results feed back into the board, follow-up tasks are created as needed.
- Questions for the user are queued in
tasks/INBOX.md— you answer asynchronously; unblocked work continues in the meantime.
git clone git@github.com:loudinthecloud/skillwave.git
cd skillwave- Open the folder in with github copilot.
- Fill in
GOAL.md— describe what you want built (use an LLM to help you if you like). - Open Copilot Chat and after seleting the
@skillwaveagent, just say "run" or "execute" to get started.
copilot --autopilot --yolo --max-autopilot-continues 10000Then select the model with /model and the agent with /agent.
| Path | Purpose |
|---|---|
GOAL.md |
Your project goal — the only file you need to fill in |
tasks/BACKLOG.md |
Task board — single source of truth for all task state |
tasks/INBOX.md |
Async questions for the user |
tasks/done/ |
Completed task reports |
.github/agents/ |
Agent definition (the skillwave orchestrator) |
.github/roles/ |
Dynamic role definitions, created on demand |
.github/skills/ |
Domain knowledge powering the orchestrator |
Orchestration — Skillwave runs a continuous loop: read the board → batch unblocked tasks → dispatch subagents in parallel → collect results → update the board → repeat.
Roles — Lightweight worker definitions (expertise, constraints, allowed files) created dynamically when a task needs a specialist. Retired when no longer needed.
Skills — Built-in knowledge modules (orchestration, delegation, task management, user comms, role creation, reporting, collaboration) that define how skillwave operates.
Task lifecycle — open → active → done (or blocked → active → done). Each completed task produces a report in tasks/done/ and a git commit.
Stateless subagents — Every worker call is a one-shot invocation with a self-contained prompt. No shared memory between agents; all coordination flows through the markdown board.
