Docs · Hangar
Autopilot: AI agent workloads on Hangar
Autopilot is Hangar’s sixth product: it runs any AI agent workload, bounded and audited, and it drives the Airframe A+ program that makes the service catalog operable by agents.
One idea: an agent is one more author of git commits. Durable changes are commits. Ephemeral runs are direct requests to Crossplane. Everything else follows from keeping those two planes apart.
Brand: the mark is Option A, Hold, chosen 2026-09-26; the five options considered are in docs/brand/autopilot-logo-options.html in the repo (open the file in a browser; TechDocs cannot link to standalone HTML), and the files are in brand/marks/.
Code: github.com/jfillman/autopilot (the AI gateway: the MCP surface, sessions, dispatch, the model proxy and the inbox; Hangar’s backends, the autopilot-gateway image, triggers, agent definitions, Preflight cases, the AppSpec planner, the agent runtimes) and github.com/jfillman/clearance (the Go core that decides and records: policy, evidence, the gates, clearance-api; D21, D22). Status (2026-11-03): M0 to M2 closed; M3’s AP-A1, AP-A2, AP-A3, AP-B1 (less commit signing), AP-B2, AP-B2c, AP-B3 and AF-8a are built and live on hangar-dev; scorecard 97.3, 14 of 14. The Clearance and Autopilot build’s W0 and W1 are Built and live on hangar-dev since 2026-11-03: gateway-v0.4.0 calls the Go core, clearance v0.29.0 (clearance-core and clearance-api from the clearance chart); v0.30.0, released, removes the Python package. Clearance is a product of its own: its runtime boundary is live, and the deployment gate’s core and its GitHub App (check runs, deployment protection rule, clearance register) are built, with admission still to come. See design.md section 0.1 and the roadmap’s M3 status.
Read in this order
| Doc | Read this when… |
|---|---|
| roadmap.md | You want the plan: milestones M0 to M5, dependencies, decisions, unverified items, and the first ten actions. |
| airframe-ai-friendly.md | You want the A+ program: the measured baseline, the ten workstreams, and the design. |
| release-file-split.md | You want to know how machine-owned and human-owned values will live in separate files, and the live proof. |
| briefings-d6-d10.md | You need to decide D6 (model hub) or D10 (when strictness starts blocking). |
| design.md | You want the Autopilot design: two planes, the workload model, tiers, policy, backends, failure modes. Section 0.1 is what Clearance becoming a product changes. |
| skyport-ai-workloads.md | You want the six demo agents, one per workload shape, and their tests. |
| glossary.md | A term is unfamiliar. |
| Diagram gallery | You want pictures: 8 plan diagrams, 19 for Autopilot, 10 reference. |
Where things are
| What | Where |
|---|---|
| Scorecard and the committed baseline | hangar/tools/airframe-scorecard/ |
Clearance: the Go core, the clearance CLI, the chart, the generated Python client |
the clearance repo (D16, D21, D22) |
| The AI gateway, Hangar backends, the gateway image, planner, agent definitions, Preflight cases, runtimes | the autopilot repo |
| Diagram generator | hangar/tools/diagrams/ |
Airframe drafts (AGENTS.md, sidecar meta) |
drafts/airframe/ |
| The Skyport demo | airframe/docs/user/skyport-demo.md and airframe/examples/skyport/ |
Status, 2026-11-03
| Design | complete; Clearance-as-a-product folded in (design.md 0.1); the 2026-11-02 build spec’s W0 and W1 moved the AI gateway into Autopilot and Clearance to Go (roadmap, “Clearance and Autopilot build: W0 and W1”) |
| Clearance core (decisions, policy bundles, the evidence chain, clearance-api) | Built, live on hangar-dev: clearance v0.29.0 from its own chart, clearance-core and clearance-api in autopilot-system; v0.30.0 released (Go only) |
| Clearance deployment gate | core Built (v0.19.0) and fed by the advisory clearance-gate- check and clearance register; admission Proposal |
Clearance GitHub App (Hangar Clearance) |
Built, live: check runs (v0.21.0), deployment protection rule (v0.22.0), clearance register (v0.23.0) |
| Evidence signing, checkpoint, witness, ledger | Proposal, on the Clearance product roadmap |
Autopilot gateway (MCP, sessions, dispatch, model proxy, inbox, person controls, run results, autopilot-api), backends, triggers, runtimes |
Built, live: gateway-v0.4.0 with its own autopilot-pg, calling the Clearance core; runtime images generic, base-python, langgraph, claude-agent-sdk |
AgentRun XRD, composition, function; Agent XRD |
Built, live on hangar-dev (airframe catalog v0.3.172) |
| Scorecard | Built; 97.3 / 100, 14 of 14 acceptance checks (baseline 27.3) |
| hangar-prod | nothing of Clearance or Autopilot runs there; Tower reads clearance-api on hangar-dev |
Run it
python3 hangar/tools/airframe-scorecard/scorecard.py # the grade
cd ~/tech/autopilot && ./.venv/bin/python -m pytest -q
cd ~/tech/clearance && go test -race ./... # the Postgres tests skip without a DSN; CI runs them