Library · Diagrams · 90 pages
Every diagram, in one place
I think in pictures, so a lot of Hangar's design lives here. Each diagram is one page with one idea, drawn with the same tool and the same visual style, and each one is clear about what's built and what's still a plan. They're generated from code in the Hangar repo, so this gallery is always up to date.
Hangar at a glance
Five ways to see the whole platform: how a change moves, the product family, where it runs, the stack and the lifecycle. They move, and every box clicks through.
Autopilot
How Hangar runs any AI agent workload: planes, shapes, runs, identity, authority, audit and rollout.
Hangar as it is, and the one new component beside it
Architecture · 02 of 19 · Two planesDurable changes are commits. Ephemeral runs are direct requests.
Secure paved road · 03 of 19 · The write pathThe durable path: an agent's write is a git write
Matrix · 04 of 19 · Workload shapesAny agent workload is one of six shapes, built from two kinds of thing
Architecture · 05 of 19 · Definition to runFrom a reviewed definition to a disposable run
Architecture · 06 of 19 · Run anatomyWhat a run gets, and the only two ways out
State machine · 07 of 19 · Run lifecycleA run always ends, whether or not anyone is watching
Tree · 08 of 19 · Agent teamsAgent teams: every child is narrower than its parent
Architecture · 09 of 19 · Backend fleetsBackend infrastructure: three fleet classes, one contract
Architecture · 10 of 19 · Model and tool planeTwo gates out of the sandbox: one for tools, one for models
Sequence · 11 of 19 · Identity and credentialsOne call, end to end: identity in, scoped token out, nothing kept
Access matrix · 12 of 19 · Tool tiersWhat each kind of agent may do, and who really does it
Architecture · 13 of 19 · AuthorityEffective authority is the minimum of five inputs, and runtime can only lower it
Flowchart · 14 of 19 · Agent PR guardrailsAn agent PR goes through the same gates, plus two that are about agents
Data flow · 15 of 19 · Flight recorderFlight recorder: reuse the telemetry, add one tamper-evident chain
Process · 16 of 19 · PreflightPreflight: score every agent change against incidents you have already solved
Architecture · 17 of 19 · AI triageAI triage: move Holmes's GitHub access behind Clearance
Deployment · 18 of 19 · Fleet placementWhere it runs: dev gets the write path and the runs, upper gets none
Gantt · 19 of 19 · RolloutRollout: four phases, each closed by a live check before the next begins
Plan
The Hangar family, the roadmap, the Airframe scorecard, the contract and the Skyport demo.
Hangar with Autopilot: six products and one reference system
Gantt · 02 of 08 · RoadmapTwenty-eight weeks, six milestones, one score to move
Dependency graph · 03 of 08 · OrderWhat blocks what: the critical path in six boxes
Matrix · 04 of 08 · ScorecardThe Airframe scorecard: ten dimensions, measured today
Architecture · 05 of 08 · The contractThe contract: one source, generated artifacts, four consumers
Sequence · 06 of 08 · One sentenceThe parachute sentence, end to end
Matrix · 07 of 08 · Skyport AISkyport gains six AI workloads, one per shape
Sequence · 08 of 08 · Event to teamA delayed flight becomes one draft, and a team, and a human decision
Reference architecture
Reference architecture for agents on a platform: substrate, gateway, identity, observability, autonomy, evaluation.
Agent execution substrate: sandboxed, scoped, disposable
Secure paved road · 02 of 10 · Tool gatewayMCP tool gateway: the only road to infrastructure
Sequence · 03 of 10 · Golden path contractGolden path as a contract: discover, invoke, verify
Flowchart · 04 of 10 · CI/CD validation loopCI/CD as a validation loop the agent can retry against
Layer stack · 05 of 10 · Multi-cloud platformOne platform surface across every cloud
Nested · 06 of 10 · Agent identityAgents are identities: each ring narrows the blast radius
Timeline · 07 of 10 · ObservabilityDid the agent do the right thing, and how do we know?
State machine · 08 of 10 · AutonomyAutonomy is earned per alert class, and revocable
Swimlane · 09 of 10 · RemediationAlerts that resolve themselves, or page with a tested hypothesis
Process · 10 of 10 · Evaluating infrastructure agentsEvaluating infrastructure agents: a regression suite, not a vibe
Glidepath
Glidepath's CI/CD architecture: system overview, chaining, multi-cluster, onboarding, deploy versus release.
Glidepath end to end: from a push to a verified release
Flowchart · 02 of 07 · What you changeWhat you change, and what the platform owns
Mapping · 03 of 07 · The contractcicd.yaml, and what actually runs
Process · 04 of 07 · OnboardingGetting a new app onto the platform
Sequence · 05 of 07 · ChainingHow test gets started when build finishes
Swimlane · 06 of 07 · Deploy and releaseDeploy and release: different questions, different rigor
Deployment · 07 of 07 · Multi-clusterOne dev control plane, and an ArgoCD in every cluster
Crossplane journey
What building Airframe on Crossplane has taught me: secret stores, readiness, API budgets and the lessons in order.
The SecretStore I planned, and the one I built
Crossplane · 02 of 04 · ReadinessThree layers, three ideas of "ready"
Crossplane · 03 of 04 · Pollers and loopsThree ways to empty a 5,000-call bucket
Crossplane · 04 of 04 · The journey so farSix weeks building Airframe, one lesson at a time
Service catalog
Why a platform needs a service catalog: one API for people, pipelines and agents, what an entry carries, and how I'd build one, discovery first.
One API, three kinds of caller
Service catalog · 02 of 05 · Two catalogsThe menu and the inventory
Service catalog · 03 of 05 · The contractWhat one catalog entry has to carry
Service catalog · 04 of 05 · DiscoveryFrom an interview to an API
Service catalog · 05 of 05 · The approachHow I'd build a company's service catalog
Beyond the cluster
Growing the catalog past Kubernetes: one API in front of many places to run, where architecture and security policy lives as code, and what a LambdaFunction does today and next.
Paving the cloud
Paving the cloud: Kubernetes and cloud resources behind one contract, what a paved road to a managed database is made of, and three weeks of the wish list from the last essay.
Any forge, any registry
Any forge, any registry: where GitHub is load-bearing today, two Hangar contracts with a driver per vendor, a short-lived token on every forge, one release on GitLab, moving 541 resources without recreating a repo, and the roadmap. A plan, not built.
Where GitHub is load-bearing today
Portability · 02 of 06 · The shapeTwo contracts, many drivers
Portability · 03 of 06 · CredentialsA short-lived token on every forge
Portability · 04 of 06 · The proofOne release, start to finish, on GitLab
Portability · 05 of 06 · MigrationMoving 88 live resources without recreating a repo
Portability · 06 of 06 · RoadmapTwenty-four weeks, GitHub first, then GitLab, then Bitbucket
Testing
Why Hangar runs its tests in Testkube: how a TestWorkflow runs, one test run through Glidepath and its policy gate, and what the free tier taught me.
Break-glass
Breaking the glass: why a debug container beats exec, how a Tower session borrows a narrow role and gives it back, two people on Flight, seven layers of guardrails, the six admission policies, the clocks, and an honest report card.
Exec, or a debug container?
Break-glass · 02 of 07 · The shapeHow a session works
Break-glass · 03 of 07 · FlightTwo people on Flight
Break-glass · 04 of 07 · GuardrailsSeven layers, each assuming the last one failed
Break-glass · 05 of 07 · The policiesSix admission policies, one per thing that could go wrong
Break-glass · 06 of 07 · TimeThe life of a session, and how it ends
Break-glass · 07 of 07 · Honest notesAn honest report card
Release record
The release record: one release from its draft to its outcome and where each step is written in git, the six questions a record answers, and the assumptions it rests on with how I would test them.
A CI/CD engine you can call
A CI/CD engine you can call: one front door to Glidepath for Tower, agents, scripts and people, how glidepath-api decides every call, and the audit line, metrics and trace one call leaves.
SLOs
How an SLO goes from a values file to a notification: the pipeline, the three signals and their thresholds, the stored state, and one poll tick.