AI-DLC, the AI-driven development lifecycle
AI-DLC is AWS’s methodology for building software with AI agents as first-class participants, published in 2025 with working materials at github.com/awslabs/aidlc-workflows. Its starting observation: when agents can write most of the code, the bottleneck moves. Typing was never the constraint worth optimising — deciding what to build, precisely enough that someone (or something) can build it, is. So AI-DLC restructures the lifecycle around fast, intense collaboration between a whole team and an AI, instead of tickets flowing through individuals.
The three phases
Inception turns a business intent into units of work. The AI interrogates the intent, the team answers, and the output is a set of units precise enough for an agent to build — each with acceptance criteria and explicit assumptions. Mob Elaboration is the central ritual here; this explainer covers it in detail.
Construction is where agents build the units. A unit labelled autonomous should be buildable from the repository and the recorded answers in one working session, with a human able to verify it in under an hour. Humans review, steer, and handle everything the elaboration marked needs-human.
Operations closes the loop: deploying, observing, and feeding what production teaches back into the next intent.
The vocabulary
- Intent — a short written statement of a business outcome: who it is for, how success is measured, what is explicitly not being done, and what mandate the team has. The unit of planning; deliberately smaller than a project and larger than a ticket.
- Unit of work — the decomposition target. One goal, testable acceptance criteria, known dependencies, and a label: autonomous or needs-human.
- Mob rituals — whole-team working sessions with the AI, borrowed in spirit from mob programming. Elaboration is the Inception ritual; the same pattern recurs during Construction.
- Bolts — AI-DLC’s name for the short bursts of agent work that replace sprints as the rhythm of Construction: hours to days, not weeks.
What changes for the team
The day-to-day inversion is that the AI proposes and the humans dispose. The agent drafts the questions, the decomposition, and the code; the team’s job shifts to answering precisely, deciding with mandate, and reviewing with attention. That only works when the inputs are honest — which is why the elaboration ritual spends so much effort separating decisions from guesses and logging what nobody in the room can decide.
It also changes what is worth measuring. Velocity of typing stops being interesting. What matters is the don’t-know rate of a question round, how many decisions leave the room without an owner, how many units survive review unchanged, and the elapsed time from intent to merged pull request.
Standing on it, carefully
AI-DLC is young, and most public material about it is description, not evidence. This site is an unofficial facilitation layer on the Inception phase: a session runner prompt, a runbook, and a growing set of pilot session metrics — because a methodology that reorganises how teams plan work should have to show numbers, not slides. If you run a session, log it.