Skip to content

Agent factory · Suite

Hand work to cloud coding agents and keep every change reviewable

Tools that let a team of six engineers run about sixty agent sessions at once. Four planes: the team harness defines what the agents run, hangar runs them, flightdeck judges the runs and Tower governs risky decisions. ai-dlc sets up the people and machines around them, and every change still leaves through a reviewed pull request.

See how it fits Open the deck

hangar's Fleet: the decision deck with a held push on top. Needs approval, Push the refunds migration to main, git push origin main, Deny, Allow for this turn and Approve push hangar's Fleet: the decision deck with a held push on top. Needs approval, Push the refunds migration to main, git push origin main, Deny, Allow for this turn and Approve push
hangar's Fleet: one decision at a time. A push to main waits for you; the agents keep working.
DefineTeam harnessThe harness as native files and a lock, frozen from a good run and promoted on evidence. RunhangarRuns coding agents in isolated cloud sandboxes, from a ticket to a landed PR. JudgeflightdeckJudges harness changes with paired statistics and behavior specs over every run. GovernTowerOne decision point for risky actions: rules, human approvals, signed receipts. Setupai-dlcSets up engineers' machines, authors the optional workflow plugin, gates local work on CI.

Two more parts round out the suite. Spaceport is the Capability: reviewed data products that agents query through four MCP tools (frozen). Airframe is Design: the system every tool uses.

See each tool work

Each clip is under 20 seconds and starts on the product. Select one to play it.

The first frame of hangar's clip: the Fleet, with issue #42's job starting in checkout-python and the caption Issue #42 gets the hangar label, and a job starts The first frame of hangar's clip: the Fleet, with issue #42's job starting in checkout-python and the caption Issue #42 gets the hangar label, and a job starts
hangar: a ticket lands
The first frame of Tower's clip: Marcus's Desk with Claude Code's git push origin main waiting for approval, and the caption Claude Code wants to push to main. Tower holds it for a tech lead The first frame of Tower's clip: Marcus's Desk with Claude Code's git push origin main waiting for approval, and the caption Claude Code wants to push to main. Tower holds it for a tech lead
Tower: a deny, with a reason
The first frame of flightdeck's clip: flightdeck stats compare in a terminal, with the two arms' resolved rates and the paired difference The first frame of flightdeck's clip: flightdeck stats compare in a terminal, with the two arms' resolved rates and the paired difference
flightdeck: a verdict on the PR
The first frame of ai-dlc's clip: ai-dlc project init in a terminal, answering applied, 37 files
ai-dlc: a project's checks pass

Start here

  • Engineer: hangar. hangar connect adds hangar to your agent client (connecting clients).
  • Harness owner: Team harness. Copy the example, freeze a good run, and promote on evidence.
  • Running it at work: Run the factory at your company. The four-repo convention, change sets from a ticket to a deploy, and what stays yours.
  • Building UI: Airframe. Tokens, components and the rules every tool's UI follows.

The four planes

The agent factory at a glanceDefine · Run · Judge · GovernThe team harness repoDefineNative files and harness.lock, frozen from a good hangar runOne identity everywhere: harness_digestai-dlc contributes the optional workflow plugin and machine setuphangarRunDoors in, admission, intake posture, taint trackingSandboxes, sessions, jobs and experiment trials on GKESealed graders; infrastructure errors retried and excludedTrusted verify, gateway push, landing on approvalflightdeckJudgeReads runs from a Parquet data plane, never from live hangarPaired statistics, behavior specs, nightly ADR checksReports and promotion evidence posted on PRsMany doorsDoorsJira “Ready for agent”GitHub label or Projects columnSlack @hangar · MCP from laptops · schedulesTowerGovernCedar rules and human approvalsThe approver is never the requesterPassports and signed receiptsFails closed: off, shadow, enforceGitHubOne gate: the PRTainted work opens as a draftA second human approves the exact SHAQuestions and reviews reach people in SlackCapabilitiesMCP servers bound into sandboxes when they are claimed: Spaceport data products, internal tools, docs searchthaw: pin the harness,check drift on turn 1outboxhangar.run/v1trialspin sets, experimentsreport, then promote the stable channelwork itemsdecidereceiptpush,open PRtools over MCP
The four planes, the doors and the gate · full size
The agent factory at a glanceDefine · Run · Judge · GovernThe team harness repoDefineNative files and harness.lock,frozen from a good hangar runOne identity: harness_digestpinned as harness.lockhangarRunWork arrives through the doors: Jira,GitHub, Slack, MCP and schedulesSandboxes, sessions, jobs and trialsVerify, push through the gatewaydecide, then a receiptTowerGovernRules and human approvals; theapprover is never the requesterSigned receipts; fails closedpush one commit, open the PRGitHubOne gate: the PRTainted work opens as a draftA second human approves the SHAevery run, through the outboxflightdeckJudgeReads runs from a Parquet data planePaired statistics, behavior specsReports and evidence on PRsreport, then promote the next harnesstools over MCP, bound into each sandboxSpaceport and MCP serversCapabilitiesData products, internal tools and docs search
The four planes, one above the other · the wide drawing

Define is the team harness repo, Run is hangar, Judge is flightdeck and Govern is Tower. hangar sits in the middle. Work arrives through the doors and the harness arrives as a lock file. Risky decisions go out to Tower, every run goes out to flightdeck, and the verdict comes back as the next promoted harness.

One number, harness_digest, identifies a harness from the lock file to the manifest, the PR and the report. ai-dlc is setup, off the agent's hot path. It prepares machines, authors the optional workflow plugin the harness may pin, and gates local work on CI.

Where to go next

  • How it fits: one ticket's journey through the four planes, the three rules, and who owns what.
  • Architecture: the system context, the containers and where they run.
  • Flows: improving the harness, a risky action, the nightly pipeline, onboarding a repo.
  • Roadmap: the four phases and their themes, with counts rebuilt daily from each repo's GitHub milestones.
  • Decisions: choices that span more than one repo.
  • Glossary: the suite's terms, defined once.

hangar, flightdeck and ai-dlc are open source. Tower and Spaceport are private repositories for now.