Skip to content

Architecture

The suite follows the C4 model. This page is the system context: who uses the suite and what it depends on. Containers is what runs where. How it fits follows one ticket through them. Both diagrams follow the page's light or dark theme; PNGs for slides are in docs/assets/diagrams.

System context

System contextC4 level 1 · who uses the suite and what it depends onPersonReviewerReviews agent PRs andapproves the exact commitPersonEngineerHands work to agents andsteers ~10 sessionsPersonHarness ownerRuns experiments andpromotes the harnessPersonApproverDecides release, landing,egress and secretsSoftware systemAgent factoryDefines, runs, judges and governs cloud coding agentshangar · flightdeck · Tower · ai-dlcEvery change leaves through a reviewed pull requestExternal systemGitHubRepos, PRs, checks, App,labels, Projects columnsExternal systemJiraTickets moved to“Ready for agent”External systemSlackRequests and questions,approval cards, digestsExternal systemModel providersAnthropic, OpenAI, GoogleAPIs, with API keysExternal systemGoogle CloudGKE, Cloud Storage,Cloud Run, registryExternal systemMCP serversSpaceport data products,internal tools, docsapproves PRshands off work,answers questionspromotesharnessesdecides risky actionspushes, opens PRswebhookswebhooksquestionsrequestsmodel callsruns oncalls tools
C4 level 1 · people, the suite and the systems around it · full size

Four kinds of people use the suite. Engineers hand off work and answer questions. Reviewers approve the exact commit that lands. The harness owner runs experiments and promotes the team harness. Approvers decide risky actions.

The suite depends on five things outside it. GitHub for code and review, Jira and Slack as doors, model providers through API keys only. Google Cloud to run on, and MCP servers for the tools agents call.

The planes

Plane Tool Decides Never does
Define The team harness repo; ai-dlc contributes the optional workflow plugin and machine setup What a harness is: native files and harness.lock, frozen from a good hangar run; one digest Run anything
Run hangar Where work runs and how it reaches a branch Analyse runs or set approval policy
Judge flightdeck Whether a change made agents better, with intervals Run sandboxes or hold credentials
Govern Tower Whether a risky action may happen Enforce inside a sandbox
Capability MCP servers, Spaceport What data and tools agents can reach Decide policy for agents