Autonomous (headless)

The agent runs unattended against a goal you defined up front. Typically long-lived monitors, scheduled jobs, or remediation workers — anything where no operator is at the keyboard. Budgets (turns, tokens, cost, wallclock) bound the run; the session log is the audit trail.

In this section

• Quickstart — first 15 minutes: a working monitor with budgets, headless permission posture, durable session log.
• Operations — the depth reference: budgets, lifecycle tool, crash-resume, failure policy, audit log, subagent composition. (Was autonomous.md pre-v2.)
• GKE multi-agent scenario — multi-agent worked example based on gke-labs/kube-agents: a parent platform agent plus operator + devteam specialists, each with their own MCP servers and skills, interacting around a real GKE rollout/SLO scenario.

Common references

• Sessions and event log — durable session storage + crash-resume
• Context management — compaction + checkpoints make long unattended runs viable
• Attach mode — let an operator drop into an unattended agent mid-run
• Agent design — prescriptive patterns; goal-following is stricter without an operator in the loop