AI Agent
A conversational assistant with read-only inspection of runs, logs, commits, costs, and runners — plus confirmed re-run actions.
The AI agent answers questions about your CI/CD estate in plain English and digs into the actual data instead of guessing.
What it can do
| Ask about | The agent inspects |
|---|---|
| "Why did the deploy fail last night?" | Error logs, stack traces, git history, diffs and PR descriptions (root cause analysis) |
| "Has this error happened before?" | Past incidents and team resolution playbooks |
| "Which repos cost the most this month?" | Compute spend and cost drivers |
| "Are we queueing on runners?" | Runner group utilization and queue congestion |
| "Show runs of release.yml on staging" | Workflow runs, jobs, deployments, schedules |
It can also re-run failed jobs or dispatch workflows — but only after asking you to confirm, so nothing executes by surprise.
Conversations
Sessions are persistent with a sidebar history, and answers render Markdown, diffs, and JSON inline.
Automatic PR failure comments
The agent's root-cause engine also works proactively: with AI Failure Summary Comments enabled in Settings → System, every workflow failure tied to a pull request posts an AI summary right on the PR — the likely root cause and suggested fix, a collapsible excerpt of the failing log lines, similar resolutions from your knowledge base, and links to the full log. The same comment updates on later failures (no duplicates), turns yellow once a failure looks flaky (5+ occurrences), and secrets are redacted before anything is posted.
Anyone can trigger a deeper pass by commenting /rca on the PR — the agent runs a full Deep RCA per failed workflow and replies in the thread. Both features are paid-tier only and count against the daily AI quota.
Daily quota
Each organization gets a daily chat allowance by plan — Basic: 50, Professional: 200, Enterprise: 1000 conversations per day. The Free plan does not include the AI agent. A usage tracker in the page shows what remains today.
The agent is read-only by design apart from explicitly confirmed run actions. It never edits workflow files, settings, or billing data.