Mob Elaboration: The Ritual That Replaces Sprint Planning

Sprint planning asks what the team can do in two weeks. Mob Elaboration asks the team to decide, in one sitting, everything the agent would otherwise guess.

Mob Elaboration is the requirements ritual of AI-DLC. The people who can make decisions about a piece of work sit in one session with an AI agent. The agent proposes the breakdown and asks the questions. The team answers, corrects, and approves. Nobody writes code. What comes out is the spec the agent will build from: requirements, user stories, units of work, and the architecture decisions behind them, all written to files in the repository. In AI-DLC terms, it is the heart of Inception, the phase where requirements and design get decided before Construction writes any code.

It replaces sprint planning and backlog refinement in the AI-Driven Development Life Cycle. The AWS whitepaper describes it as one room, one shared screen, one facilitator, and an AI that proposes user stories, acceptance criteria, and Units while the mob fixes what came out under-engineered or over-engineered. The paper’s claim is that it condenses weeks or months of sequential work into a few hours.

That claim is true, and it is also the least useful thing you can know about the ritual. The paper tells you what Mob Elaboration is. It does not tell you who should be in the room, what to do before the session, how to answer the agent’s questions, or why the third hour goes so badly. I learned those things watching teams run it inside a large engineering organization, through pilots and a multi-day training for the people who will spread it. This is the guide I wish they had been handed on day one.

Mob Elaboration compresses weeks of refinement into one sitting

Think about how a feature gets refined in a normal agile team. The PM writes a ticket. A developer has a question and asks it in a chat channel. The designer answers two days later. Someone decides something in a hallway, and the decision lives in one person’s head until it surfaces as a bug. Refinement is not slow because anyone is lazy. It is slow because it happens in fragments, across days, with each fragment waiting on a person.

With an agent that can generate a full set of requirements in minutes, that waiting becomes the whole cost. Mob Elaboration attacks the waiting directly. It puts everyone who can answer in the same session and lets the AI drive the questions in order, so a decision that used to take a week of back and forth takes five minutes of talking.

The name comes from mob programming, the practice Woody Zuill’s team made popular around 2012: the whole team works on the same thing, at the same time, at one computer. The difference matters. A programming mob writes code together. An elaboration mob decides together, and the AI does the writing.

Only people who can decide get a seat

The single rule that made the biggest difference in the sessions I watched: only someone with decision power over some dimension of the work joins. Product decides scope and business rules. Engineering decides architecture and trade-offs. Design decides the experience. Data decides what gets measured. Everyone else is an observer, and observers make the session more expensive without adding context.

That sounds harsh until you count the cost. A mob of eight for three hours is a full day of one person’s work. If three of those eight cannot answer any question the agent asks, a third of that day bought nothing.

The right mob changes with the work. The pattern that settled in:

Who sits in the mob

Three to five people is the sweet spot. Each one should be able to answer at least one kind of question the agent will ask.

Kind of work	Who joins
User-facing feature	Engineering, product, design
Data or machine learning	Engineering, data, product
Purely technical change	Engineering and an architect
Algorithmic optimization	Engineering and a domain specialist

AI-DLC 2 even has a stage for this. In the Feature and Enterprise profiles, Team Formation (stage 1.5) has the delivery agent assess availability and skills and write a mob composition plan before Inception starts. It is a good stage. It still cannot tell you that your tech lead is going on vacation next week.

Run the reverse engineering before anyone joins

If the work touches an existing codebase, the agent has to read it before it can ask good questions. AI-DLC calls this reverse engineering: the agent scans the modules the feature will touch, maps how they talk to each other, and writes down the stack, the patterns, and the architecture. Every question it asks afterwards is filtered through that model. Without it, the agent proposes changes that duplicate logic that already exists or break contracts nobody wrote down.

On a small repository that pass took about twenty minutes in the sessions I saw. On a large one it takes longer. Either way, it is twenty minutes of a mob watching a progress bar, which is the most expensive way to spend twenty minutes.

So the person driving runs it ahead of time. Install the workflow, start the intent, let reverse engineering finish, review and approve what it produced, and only then call the team. People walk in to a questions file that is ready to answer. The details of doing this on a large codebase, including how much context it eats, are in AI-DLC in brownfield.

The questions file is the meeting

After reverse engineering, the agent writes a questions file. It is a plain markdown document with numbered questions, multiple-choice options, and an empty answer tag under each one. In one session I watched, the agent produced fifteen questions for a small backoffice feature. They covered scope, behavior, business terms, and technical decisions the original ticket never mentioned.

A question looks like this:

## Question 4

How should a cancelled order affect stock that was already reserved?

A) Release the reservation immediately
B) Release it after a grace period
C) Keep it until someone confirms the cancellation
X) Other (please describe)

[Answer]:

Here is the reframe that made teams stop resenting the file. Every one of those questions is a question a developer would have hit in the middle of implementation. Without the ritual, they would have stopped coding, pinged someone, and waited. With it, the question gets answered before code exists, and sometimes the answer changes the whole approach. The file does not add work. It moves work to the cheapest point to do it.

Teams also had to unlearn treating the file like an exam. These four habits fixed most of that, and each one comes with words you can type:

How to answer the agent's questions

1. 01

   Answer outside the options when the options are wrong.

   Early on, teams felt locked into A, B, or C. The options are the agent's guesses, not constraints. It reads a free answer just as well.

   Type this

   None of these. Reservations stay until the warehouse system confirms the cancellation, because a manual refund can still reverse it.

2. 02

   Defer what is not blocking, and give it an owner.

   A question that depends on an external decision should not stall the session. The agent skips it and asks again later. Blocking questions, like an API contract the Units depend on, it will not let you skip, and it should not.

   Type this

   We don't know yet. Mark it TBD, owner: the payments lead, and ask again before Units Generation.

3. 03

   Change your mind out loud. Answers are reversible.

   The agent rebuilds the documents from the current answers. Changing an answer sends the workflow back to the stage that depended on it. You never need to kill the session and start over.

   Type this

   Change question 4 from A to B and update every document that depended on it.

4. 04

   Ask before you choose.

   When nobody in the room understands an option, ask the agent to explain it in the context of this project. The prefix stops it from treating your question as an instruction to edit.

   Type this

   Do not update any documents. What would option C mean for the current reservation flow?

One more habit is worth adopting: keep a list of the questions the agent asked that the original ticket already answered. That list tells you what is missing from the input. Feed it back into how your team writes the next intent, and the next session gets shorter. Where that input should come from is its own problem, covered in AI-DLC’s blind spot.

Product joins for the business questions, not the whole session

The calendar is the first thing that kills Mob Elaboration in a real company. Getting a PM, a designer, and three engineers into the same room for three hours, every time a feature starts, does not survive contact with anyone’s week.

The fix came from a product manager in the pilot, and it was simple. Most of the questions after reverse engineering are technical. The engineers answer those on their own. When they reach a question about business rules, they call product in. In the case they described, it took ten minutes of their time instead of an afternoon. Their line, more or less: not everyone needs to be in the room at the same time.

Product still owns two things nobody else can sign. The first is whether the questions are even relevant, because the agent sometimes asks about things the ticket already settled. The second is the success criterion, which is the single most common thing the agent gets wrong. That trap gets a full section in AI-DLC’s blind spot.

Decide how much of the spec the AI writes

One team I followed made this choice explicit at the start of every task, and I think every team should. There are three ways to split authorship of the spec between the humans and the AI:

Three ways to split spec authorship

Collaborative is the right default. Strict and Off are deliberate exceptions, chosen out loud.

Mode	Who writes	Use it when	The risk
Strict	Humans write every line. The AI only validates.	Heavily regulated or sensitive domains.	Slow. You give up most of the speed.
Collaborative	The AI asks, the team answers, the AI drafts, the team approves.	The default for almost everything.	Gates degrade when people are tired.
Off	No spec. The agent goes straight to code.	A spike you will throw away.	This is vibe coding with extra steps.

AI-DLC 2 has a related but different choice, and people mix them up. Every stage that gathers input offers three interaction modes: Guide Me (the agent asks one question at a time), Edit File (you fill the questions file yourself), and Chat (free conversation, with the agent writing the decisions back to the file). That is about how you answer. The three modes above are about who writes the spec. You can be Collaborative and answer in Edit File mode. Both converge on the same questions file.

One hour, then stop

This is the part of Mob Elaboration the documentation never mentions, and it is where most sessions fail.

The AI produces a requirements document, a set of user stories, and a list of Units in minutes. A human needs real time to read each one with care. The gap between generation speed and reading speed is wider than with any tool before, because the AI does not hand you a suggestion, it hands you a finished artifact. After about an hour, people stop reading and start approving.

An engineer in the training said it out loud near the end of a long Inception: by the time the Units of work came up for review, they had nothing left to review them with, and they knew that if they kept going they would approve anything. They stopped and came back another day. That was the right call, and most people do not make it.

Pace rules that held up

• ✓Required:
  Cap every session at about an hour.Split a large Inception across two or three sessions on different days.
• ✓Required:
  Pause at gates, never in the middle of a stage.The state file records where you stopped, so resuming costs nothing.
• ✓Required:
  Commit and push before you clear the context.The work lives in the files. The conversation is disposable.
• ✓Required:
  The manager sits in the first two or three mobs.Not to observe. To call the break when approvals start coming without questions.
• ✗Anti-pattern:
  Approve without being able to say why the artifact is right.If you cannot explain it in one sentence, you are not done reading.

The deeper version of this argument, with the other ways approval gates fail, is in gates are a loss function.

What AI-DLC 2 automates, and what it leaves to you

AI-DLC 2 moved part of the mob inside the machine. The Inception phase now runs nine stages, from reverse engineering to delivery planning, and one of them is built as a mob of agents. In User Stories (stage 2.4), the product agent drafts the personas and stories. Then the design, developer, and quality agents each review that draft in parallel, without seeing each other’s notes, and write their objections. The product agent folds them in. Judgment calls where both positions are legitimate go to you as a structured question. Factual disputes get one more bounded round between the agents. A separate product reviewer checks the result before it reaches your gate.

A Mob Elaboration in AI-DLC 2, start to finish

Input

An approved intent, and a repository the agent can read

↓

1. 2.1Reverse engineering, before the team joins

   A developer agent scans the code, an architect agent writes the synthesis. One complete pass per repository when the work spans several.

   ↓
2. 2.3The questions file

   Engineers answer the technical questions. Product joins for the business ones. Answers can be free, deferred, or changed later.

   ↓
3. 2.4User stories, drafted by a mob of agents

   The product agent drafts; design, developer, and quality agents object in parallel; you settle the judgment calls.

   ↓
4. 2.7Units of work

   The agent decomposes the stories into Units with a dependency graph. This is where you cap their size.

   ↓
5. GATEApprove or request changes, then stop

   A verification gate checks that every requirement traces to a story before Construction begins.

↓

Output

A spec in the repository that the agent can build from, and a team that agrees on what it says.

That is real progress. Agents arguing with each other before you see the draft catch a class of mistakes that used to reach the gate. But look at what the engine still cannot do. It cannot choose who sits in the room. It cannot notice that the PM has stopped reading. It cannot know that the business rule it just wrote down is the one the company abandoned last quarter. The ritual’s hard parts are human, and they stayed human.

Planning is not Mob Elaboration

Once a team adopts the ritual, the rest of its calendar changes, and teams that do not change it end up with two planning meetings.

Planning decides what to work on next. Mob Elaboration decides how one chosen piece of work will behave. They are different meetings with different people. The daily shrinks to a few minutes, because the state of every piece of work is in its files. Kanban fits better than sprints, because work now arrives in slices of hours or days, not two-week packages.

The backlog itself took one of three shapes in the teams I watched:

Three shapes of a backlog under AI-DLC

The third shape is still AI-DLC. It is just a mob of one, plus the agent.

Shape	How it runs	When it fits
One big initiative a week	The whole squad elaborates together, then builds.	A large feature with many open questions.
Several small tasks	A morning of elaboration, then construction runs asynchronously on isolated branches.	Independent tasks that each need a few decisions.
One task, solo	One experienced engineer and the agent, no mob.	Well-bounded work with no cross-team decisions.

And keep the floor in mind. One engineer in the pilot ran the full ritual to change a single line of code. If you already know exactly what to change, change it. Mob Elaboration pays when the work needs decisions from more than one person. Below that, it is ceremony.

If your team is new to writing specs at all, the session will feel like an interrogation, because every question exposes a decision nobody made. That is the method working, but it is also a sign to build the habit first, the way Spec-Driven Development for teams describes, before you run 33 stages.

FAQ

What is Mob Elaboration?+

Mob Elaboration is the requirements ritual of AI-DLC, the AI-Driven Development Life Cycle. The people who can decide about a piece of work meet with an AI agent, the agent proposes the breakdown and asks structured questions, and the team answers and approves. It produces requirements, user stories, Units of work, and architecture decisions as files in the repository. No code is written.

How long does a Mob Elaboration take?+

The whitepaper says hours instead of weeks, and that matches what I saw for a single feature. But do not run it as one long block.

After about an hour, reviewers start approving without reading. Split a large Inception into one-hour sessions, prepare reverse engineering before the team joins, and pause at gates.

Who should attend a Mob Elaboration?+

Only people with decision power over some part of the work: typically engineering and product, plus design for user-facing work, data for data work, an architect for purely technical changes, or a domain specialist for algorithmic work. Three to five people is the sweet spot. Observers make it more expensive without adding answers.

Is Mob Elaboration the same as mob programming?+

No. They share the idea of a whole group working on one thing at the same time. A programming mob writes code together at one keyboard. An elaboration mob makes decisions together while the AI does the writing, and it ends before any code exists.

Can Mob Elaboration be remote or asynchronous?+

Yes. The session runs fine on a call with a shared screen. Parts of it can be asynchronous: engineers answer the technical questions on their own and product answers the business ones when called. What cannot be skipped is that every answer ends up in the questions file.

Does Mob Elaboration replace sprint planning?+

It replaces refinement, and it changes planning. Planning still picks what to work on next. Mob Elaboration decides how that work behaves. Most teams find Kanban fits better than two-week sprints once work arrives in slices of hours or days.

Where to go next

Mob Elaboration is where AI-DLC either earns its keep or turns into theater. The agent is good at asking questions. The ritual only works if the right people answer them, at a pace that lets them actually read what they approve. Prepare the session, keep it short, and write every decision down.