Documentation
Plan

45 of 187 pages

AI Product Studio

Agents from Your Repository

Your repository may already define its coding agents — a sceptical enterprise reviewer, a bug fixer, a test author. The Studio can bring those definitions onto its roster, have them review your checklist items under their own instructions, and send them back to the repository as files.

The two formats

.github/agents/<name>.agent.md — GitHub Copilot custom agents.
.claude/agents/<name>.md — Claude Code sub-agents.

Both are markdown with a short header (name, description, tools) and a body that is the agent's instructions. A body with What I do and Guardrails sections maps onto the roster agent's goals and guardrails.

Importing

On the Agent Roster, the Sync with repository card lists the definitions in a connected GitHub repository and shows, for each, exactly what an import would put on the roster: the persona, goals, guardrails and skills. Tick the ones you want and choose Import. Each becomes a custom agent marked From repository, placed in a role by what it says it does. Import it again after the file changes and the same agent is updated.

What you reviewed is what is imported. The import re-reads each file from the repository and checks it against the version you were shown. If the file changed in between, the import is refused and the list refreshes so you can review the new version.

Importing is for studio owners and reviewers, and imported agents are custom agents, so they need a Business or Enterprise plan. Instructions are stored up to a length that keeps the studio record healthy, and skills are kept as short tags.

Naming an agent on a checklist item

On the Work Package Checklists page, a Definition of Ready or Done item can name one particular agent from the roster rather than just a role. The gate then runs that agent — not any agent in the role — and records its name on the evidence. If the named agent is later disabled, the item is handed to a person rather than quietly given to another agent.

The assessment runs as that agent: its persona and standing brief lead the instructions. What it is asked to verify, the guardrails around that, and the shape of its answer stay ours, so an agent's instructions can shape how it works but never what it is allowed to conclude. For the same reason, changing a roster agent's instructions is an owner or reviewer decision, and every change is recorded in the audit log against the agent.

Running as the agent on the pull request

When a synced agent raises a coding action on a pull request, the Studio asks the pull request's coding agent to carry it out as that agent: Copilot is pointed at the custom agent of that name, Claude at the sub-agent. On a system-runner pull request, the corrective run uses the agent's instructions directly.

Exporting

The 📤 button on a roster agent writes it to the repository in both formats. It never pushes to your default branch: it opens a pull request on a pw/agents-* branch, and it can write only inside the two agent folders. Exporting is for the studio owner, and the pull request is opened as you: connect your own GitHub account under My Account → Source control first.

All AI Product Studio documentation

Build v0.3.2 · deployed 2026-10-19T23:52:31Z