User guideAI Engineering Platform
Back to siteSite

Key concepts

The platform implements an AI Software Development Lifecycle (ASDLC): a sequence of phases, artifacts and gates. The model is replaceable; the process is the product.

Lifecycle at a glance

idea → blueprint → backlog → ready → plan → approve → build → test → review → PR approval → merge → measure

Each arrow is a state transition. The platform records who or what caused it.

Phases

PhaseOwnerOutput
Blueprinthuman + agentproject template: stack, repository layout, tests, agents
Backloghumantasks with context and acceptance criteria
Definition of readyhumana task that can safely start
Planplanning agentsteps, files, risks, estimated cost
Builddevelopment agenta diff in an isolated sandbox
Testagent + platformtest report, design checks, acceptance evidence
Reviewhuman + agentfindings, PR comments, approval or rejection
Mergehumana merged change in the Git provider
Measureplatformdelivery and quality metrics

Gates

A gate is a deliberate stop. It cannot be skipped silently; an explicit override is recorded as an event.

  1. Plan approval — a human accepts the plan before implementation.
  2. Test gate — required checks must pass; acceptance evidence must be valid.
  3. Design gate — design standards and visual checks for UI changes.
  4. PR approval — a human approves the pull request.
  5. Merge queue — changes land in a serialised order.

Artifacts

Every artifact is versioned and hashed:

  • context manifest — the exact files, rules and token budget given to the agent
  • plan — steps, files, risks and cost estimate
  • diff — the change produced by the run
  • test report — unit, integration, e2e and visual results
  • review — findings and their severity
  • audit event — append-only record of a decision or transition

Roles

RoleTypical responsibilities
Ownerworkspace billing, members, security settings
Adminprojects, providers, rules, integrations, audit
Membertasks, runs, reviews, pull requests
Viewerread-only access to projects and results

Roles are enforced by the API and the UI; project-level access can be narrower than the workspace role.

Runs

A run is one agent execution with a type: planning, development or review. Runs are resumable at phase boundaries and stop at approvals. A run has a budget, a frozen context and an append-only event stream.

Two kinds of speed

The platform reports delivery speed and outcome quality:

  • lead time and approval time
  • first-pass merge and proof validity
  • rework, change failure rate and escaped defects

A faster merge that creates more rework is not an improvement.

Next