autonomous-setup skill

The autonomous-setup skill walks a user through configuring an unattended core-agent — single-agent monitor or multi-agent team. Bundled in SKILLS/autonomous-setup/.

What it covers

A 9-step runbook:

1. Decide if autonomous actually fits (vs interactive + attach)
2. Define the goal crisply
3. Pick single-agent or multi-agent shape
4. Bound the run with budget caps
5. Write AGENTS.md with unattended discipline
6. Set headless permission posture
7. Wire durable sessions
8. Write the Go driver
9. Test end-to-end before deployment

The runbook is non-linear; the agent picks references based on which path the user is on.

Triggers

The skill matches phrasings like:

• “set up an autonomous agent”
• “run core-agent unattended”
• “configure a monitor”
• “build a multi-agent team”
• “set up a long-running agent”
• “watch the deploy / SLO / queue”

Any phrasing implying unattended execution OR mentioning concrete monitoring/coordination tasks.

References

Three reference files; the agent fetches based on the user’s path:

• single-agent-monitor.md — single-agent monitor pattern with a full worked example (Kubernetes deployment watcher). Covers goal-sentence shape, AGENTS.md, config.json, the Go driver, attach-mode observation, variations (cadence, multi-cluster), common failure modes. Read for ~80% of autonomous setups.
• multi-agent-decomposition.md — parent + specialists pattern (from gke-labs/kube-agents). Covers when to split, scope separation via tool surface, parent / specialist AGENTS.md patterns, cost model, anti-patterns. Read when the workload genuinely decomposes (multi-tenant, multi-domain).
• budgets-and-recovery.md — sizing the four budget caps, per-agent vs shared budgets, failure policy, crash-resume via ResumeAutonomous, audit-log queries, diagnosing runaway cost. Read when sizing or troubleshooting.

When to invoke vs read the docs

The skill IS the docs in workflow form.

Installing

cp -r /path/to/core-agent/SKILLS/autonomous-setup .agents/skills/

See Skills library → Installing for global install options.

Adapting

Common adaptations:

Org-specific budgets. Edit the skill body’s “picking values” section to hardcode your org’s defaults:

For all production agents at Acme, use these defaults:
- WithMaxWallclock(24*time.Hour)
- WithMaxCost(20.00)
- WithPerTurnTimeout(60*time.Second)
Cost overruns require approval; smaller budgets are encouraged.

Pre-defined templates. Replace the generic Go driver with one that points at your team’s template repository:

For Acme autonomous monitors, start from our template:
git clone github.com/acme/agent-template my-monitor
Then customize cmd/main.go's goal string.

Cloud-specific patterns. Add a references/aws-deployment-patterns.md (or GCP, Azure) covering your cloud’s runtime quirks — service account setup, log forwarding, etc.

Where to go next

• Autonomous quickstart — single-agent worked example
• GKE multi-agent scenario — multi-agent worked example
• Subagents and wrappers — choreography patterns the skill references
• Cost efficiency — model-selection decisions for autonomous runs