7. Use the Autoheal Assistant
The Assistant is Autoheal's own agent. It knows the agent format, reads your project, and writes changes for you to review. In this chapter, you build a second agent from a description.
Step 1: Describe the agent
Go to Agent Studio → Agents and click Create agent. Type release-notes as the agent name. In What should it do?, type:
When a release is published, write release notes from the pull requests
merged since the previous release. Group them under Features, Fixes and
Internal. Read-only GitHub.
Click Create with AI. If the dialog has no Create with AI button, the Assistant is not available in your organization. Build the agent as in chapter 1 instead.
Step 2: Review what the Assistant writes
The editor opens on the new agent, and the Assistant panel opens next to it. The Assistant writes agent.yaml and instructions.md as a draft. Its changes show in the editor's Changes panel. Read the files. To change them, ask the Assistant in its panel, for example: "Use claude-sonnet-5, and put breaking changes first."
Step 3: Commit
When the files are correct, click Commit changes in the Changes panel. If the Assistant commits a change itself, it first shows an approval card, and it commits only after you approve.
Step 4: Ask for a trigger
Click Assistant in the top bar, or press ⌘J (Ctrl+J on Windows and Linux), on any page. Type:
Add a trigger that runs Release Notes when a release is published in acme/web.
The Assistant tests the trigger against a sample event and shows you the result. It saves the trigger only after you approve.
What the Assistant can do
- Create and change agents, skills, evals and triggers.
- Start runs, follow them, and explain them: "Why did my last PR Review run say do not merge?"
- Read the Engineering Context Graph.
- Set up a project's context, so that its agents can investigate. See Onboard a project.
- Tell you how to connect an integration.
- Answer questions from these docs.
It asks before it commits a change or saves a trigger. It cannot delete triggers.
Your conversations with the Assistant belong to you, not to a project. One conversation can work across projects.
Onboard a project
A project usually holds one domain: one business area and the team that runs it, for example a payments platform. Before the project's agents can investigate an incident, they must know what the domain does, which components it has, and where each component's logs, metrics, alerts, tickets and code are. The Assistant can collect this knowledge with you.
Open the Assistant and type Set up context for my project (or Review my setup to continue an earlier session). The Assistant then does these steps:
- It asks which project to set up, and finds the project's team in the Context Graph.
- It checks which integrations are connected and that they work. It gives you a connect link for each tool that is missing. It reads the integration skills, context forms and guides that already exist, and it can import a skill from a trial project.
- It reads the wiki pages and docs that you link, or that a wiki integration finds.
- It asks about one topic at a time: the domain and its components, alert sources, logs, metrics, cloud resources, deploys, data stores, tickets, chat channels and past incidents. For each topic, it first looks in your integrations and shows you what it found. Then you confirm or correct it.
- It lists your repositories, asks which ones to read, and reads them: entry points, configuration, log messages and deploy files.
- It writes what it learned as skills in the project: an architecture skill, a skill for each important component, a section for your team in each integration skill, a codebase skill, and, if they apply, a data skill and an incident-patterns skill. It runs each query on your integration before it writes the query, and marks a query that it could not run as untested.
- It adds your team, its services and their owner, and
teamlabels on your repositories to the Context Graph. - It ends with a readiness report: what is covered, what is missing, the open questions for your team, and a past incident to try with your triage agent.
The Assistant shows each skill before it commits it, and commits only after you approve. It writes Context Graph entries without an approval card, so it asks you to confirm the list first. If several people onboard the same project, the Assistant adds to the existing skills and does not replace them. You can stop at any time: what the Assistant wrote stays, and Review my setup continues from the gaps.