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. AI drafts plans, asks for clarification and proposes implementation; people validate the direction and make consequential decisions. Faster code generation does not remove the need for clear requirements, testing, security review or operational ownership.
The three phases in AWS's original explanation
This overview follows AWS's introductory article. Workflow implementations evolve and adapt their stages to the task.
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 in this site's runbook 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 working pattern is that AI proposes and people decide. 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.
How AI-DLC differs from Scrum and SAFe
These are not interchangeable frameworks. Scrum provides a framework for developing products; SAFe addresses coordination at scale. AI-DLC focuses on AI-assisted delivery workflows. You can trial this site's elaboration session without replacing your existing framework:
- Make units reviewable. This runbook aims for work an agent can attempt in one session and a person can review in under an hour. That is a local sizing guideline, not a universal AI-DLC rule or a limit imposed by Scrum.
- Keep forecasting where it helps. Record open decisions and compare actual delivery outcomes. AI assistance does not remove uncertainty or the need to plan capacity.
- Timebox the first pass. The planned workshop takes at most 2 hours 25 minutes. Revisit scope when implementation, feedback or new evidence changes what the team knows.
- Make context accessible. This runbook saves artifacts in the repository. A connected tracker can also provide context; agree which record is authoritative and keep it current.
Whether this helps your team is an empirical question. Compare a similar piece of work, include human review time, and record rework rather than treating faster drafts as proof of faster delivery.
How to start without rewriting your process
The cheapest entry point is a single Inception ritual on one real intent, leaving everything downstream untouched. Write a half-page intent, run one Mob Elaboration session with the session runner prompt, and let your existing sprint process consume the resulting units as it would any other backlog items. You keep your tracker, your standups and your release process; the only thing that changes is how scope gets written.
Two failure modes to avoid on the first attempt: adopting the vocabulary without the discipline (calling stories “units” while acceptance criteria stay vague), and treating the AI as a decision maker rather than a process runner. The runbook is written mostly to prevent these two.
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 place to share feedback. Public session logging is paused. Keep your own baseline and post-build observations; comments alone do not establish a productivity improvement.