← Back to blog

Sprint Length With AI Agents: Why Ours Is Two Days

Sprint length with AI agents should follow how often a team can decide. How a small team runs two-day cycles, and when two weeks still fits.

Our sprints used to last two weeks. Now they last two days, because the three of us meet every other day and a sprint is whatever happens between two calls.

That change came from how we already worked. At YAGNI we build AI agents for engineering teams, and we build them with our own agents. Writing code stopped being the slow part of our week a while ago. Deciding what to build next stayed slow, and that is the part the short cycle addresses.

Search for sprint length and AI and you find an argument about whether to move to one-week sprints or keep two weeks and run several review loops inside each one. Both sides measure the sprint against coding speed. This post measures it against decision speed, which is what agents changed for us, and covers what a two-day cycle looks like, what breaks, and when two weeks is still the better choice.

Why were sprints two weeks in the first place?

The Scrum Guide says a Sprint is a fixed length of one month or less. Two weeks became the common choice because it balanced two costs. A sprint had to be long enough for a team to build something worth reviewing, and short enough that a wrong direction got caught before it ate a month.

Most of those two weeks went to building. Planning took a morning, review and retrospective took an afternoon, and everything in between was engineers turning tickets into code. The sprint length was set by how long the building took.

The ceremonies around it existed to make decisions with evidence. Planning decided what to build, review checked what got built, and the retrospective decided what to change. The two weeks gave each decision enough new information to be worth a meeting.

What changes when agents do most of the building?

When agents carry a large share of the building, the time a ticket spends in code drops, and the two weeks stop being about building. Tickets get done early in the sprint, and the rest of it is waiting: for the next planning session, for someone to decide what is next, for a review queue that a person still owns.

The work that stays with people is deciding. Which ticket matters now, whether the plan is right, whether the pull request should merge, and whether the direction still makes sense after what a customer said yesterday. Agents make the building cheaper, and they make it cheaper to try a direction and drop it.

That changes the question. The useful sprint length is the interval at which your team has enough new evidence to change its mind. With people writing all the code, that interval was about two weeks. With agents writing much of it, for a small team, it can be a few days.

How short can a cycle get?

As short as the time it takes to produce something worth deciding about. For us that is two days.

We are three people. Every other day we get on a call, go through what happened since the last one, and decide what matters most now. Each of us leaves with what we own until the next call. Some cycles that is engineering. The next one can be sales, or positioning, or whatever is blocking the business that week. We can change direction two or three times in a week, and the company keeps running.

Two conditions make it work. The first is that two days of work produces real evidence: merged pull requests, a run that failed, a reply from a customer. If a cycle ends with nothing to look at, the call has nothing to decide on and should not happen. The second is that the facts are already written down. If the first twenty minutes of a call go to people reporting what they did, a two-day cadence costs more than it gives.

What does a two-day cycle look like?

Here is one cycle on a team working from a Linear backlog with agents and YAGNI’s reviews, laid out against the parts of a two-week sprint they replace.

Two-week sprint Two-day cycle
Sprint planning picks tickets for two weeks The call picks what matters for the next two days and writes it into the tracker as tickets
Engineers build through the sprint An agent builds each ticket to a pull request
Code review happens when someone has time Proctor reviews every pull request when it opens; by default a person merges
Daily standup reports status The call reads what merged and what failed straight from GitHub and the tracker
Sprint review and retrospective at the end The next call starts from what merged, what failed, and what customers said

The call itself has three parts. First, what happened: what merged, what broke, what a customer or a pilot team said. Second, what matters now, given that. Third, who owns what until the next call. The decisions go into the tracker as tickets or changes to existing ones, so the Team works from the same place the people do.

People decide what gets built. The call picks the work, and a person owns each ticket the agents pick up. Proctor reviews every pull request, and by default a person merges. A shorter cycle means people make those calls more often, and they stay people’s calls. For how the record of each decision works, see how to track engineering progress without a project manager.

What breaks when you change direction every two days?

Customers feel every turn. A pilot team that meets once a week can see the product change two or three times between its own meetings, and some of those changes reverse earlier ones. We now show customers only what we trust and keep experimental work hidden until it holds. We also tell them, in plain words, which parts are stable and which are still moving.

Tickets go stale faster. A ticket written under Monday’s direction can be wrong by Wednesday. The fix is to keep the backlog honest at every call: close what the new direction made unnecessary, and rewrite what changed. A person decides what no longer matters.

Focus is harder to protect. When direction can change every two days, it is tempting to change it every two days. The calls need a longer goal to measure each turn against, or the team ends up reacting to whatever happened most recently. We keep strategic goals that span months, and the two-day cycle decides the next step toward them.

Status reporting becomes a tax. Three calls a week of people reading out what they did would cost more than the cadence saves. The record has to come from the systems that already hold it, which is the same argument behind an async standup built from GitHub and Jira.

When should you keep two-week sprints?

A two-day cycle fits a small team where everyone is on every call and nobody needs a summary of what was decided. It gets harder as the team grows.

Keep a longer sprint when several teams depend on each other’s work and need a shared rhythm to plan around. Keep it when releases ship on a fixed train and changing direction mid-train is expensive. Keep it when most of the cycle is still people building, because then the building time still sets the pace. And keep it when the decisions themselves need more than two days of evidence, such as a pricing change that needs a full billing period of data.

A team can also mix both. Some teams run a longer planning rhythm for the roadmap and short cycles inside it for what the agents work on next. The sprint length that matters is the one at which your team decides.

Where to start

Try it for two weeks. Meet every other day, keep each call to what happened, what matters now, and who owns what, and stop the experiment if the calls run out of things to decide. If your team already works with agents, the evidence for each call is mostly in GitHub and your tracker.

Keep three questions in front of you while you try it. Did each call have something new to decide on, or did it repeat the last one? Did a decision from one call get acted on before the next, or did it wait? And did anyone spend the call reading out status the tracker already had? If the answers are mostly yes, no, and yes, the cycle is too short for your team right now, and a weekly rhythm is a better place to start. If the calls keep finding real decisions, the old sprint length was holding them back.

Whatever length you land on, write it down as a decision with the reason beside it, and look at it again when the team or the work changes.

For more on working with agents day to day, see AI agents for software development and managing AI coding agents. If you want to see YAGNI work a ticket from your backlog, book 30 minutes.