Anatomy of one idea: the ADLC from first sentence to counted delivery

The best way to explain the Agent Development Life Cycle is to replay a real idea. Idea #123 in our own workspace — "Track AI-written lines of code per idea" — went from one sentence to delivered feature with every step recorded on the idea itself. Here is its actual timeline.

Station 1 — the idea documents itself

The idea was filed with a short description of the problem. Within seconds, UltiPad drafted three user stories with acceptance criteria ("When I review my list of ideas, I want the AI-written line count visible for each idea, so I can compare productivity across ideas") and a functional spec covering the data model, the tool changes, and the reporting surfaces. Nobody wrote a PRD. The meeting that would have produced one never happened.

Station 2 — the agent spec is the contract

A human approved the approach — one decision, made once: line counts would be agent-reported from git's own diff stats rather than fetched from a provider API. That decision became the agent spec: machine-readable, human-approved, and the exact contract the coding agent builds against. This is the trust boundary of the ADLC — agents never act on vibes, only on approved specs.

Station 3 — build, and the wire runs both ways

The agent pulled the spec over MCP and reported "started" — which moved the idea's workflow stage to In Progress automatically. Then came the commits, each recorded against the idea with its diff stats: +396/−11 across 12 files for the core feature, +23/−6 for a race-condition fix, +62/−1 for webhook auto-detection. Total: 481 AI-written lines on this idea, visible as a summary strip on its Commits tab.

That middle commit is worth pausing on. During rollout we discovered the git webhook usually records a commit before the agent can claim it — stealing the attribution. The agent filed the problem, fixed it (an agent re-claiming a webhook-linked commit upgrades its attribution and back-fills the stats), and shipped it through the same loop. The lifecycle debugged itself.

Station 4 — delivery is a state, not a status report

When the agent reported "completed", the idea moved itself to QA for human verification, the spec status updated, and the delivery rolled up into the reports: the funnel counts it, the AI-contribution chart adds its 481 lines, and the owner's delivery scorecard credits it. Every claim in this article is checkable on the idea's own page — which is the point of the ADLC: the audit trail is the workflow.