Working paper · v0.1
A sprint methodology for AI agent teams, with observations from 324 sprints
Dave Remy · 28 October 2026 · doi:10.5281/zenodo.23771171
324
two-hour sprints, run around the clock, 20 Sep to 28 Oct 2026
79%
of committed stories done inside the sprint (297 of 375, team sprints)
<5 min
of ceremony per sprint: starting plus wrapping
<2 h
a day of human time, usually under one hour
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171
Scrum was not the plan. The first sprint came from a request for sessions that end in a summary and a retrospective. Every later device was adopted reluctantly, one at a time, to answer a specific failure. The boundary does three jobs that an open queue tends to postpone:
Write a durable handoff and start fresh on a schedule, before the context fills and quality drops.
Unfinished work is saved and accounted for, not left to pile up silently.
Committed against done, every two hours, which is what makes velocity and planning possible.
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171
| Scrum | Agent Scrum |
|---|---|
| Stakeholder | The human: direction, judgment, taste |
| Product owner | The CTO agent: backlog, planning, rulings; runs no sprints |
| Scrum master | A scrum-team agent, started on demand, one sprint at a time |
| Developers | Short-lived developer and reviewer agents in isolated worktrees |
| Sprint length | Two hours, on fixed boxes starting at odd hours |
| Velocity | Rolling merged points per sprint, plus token velocity |
| Daily stand-up | None: a live board instead |
| Retrospective | Every wrap; each finding ends in an issue, a fix or a reason |
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171
Human
Stakeholder: direction, taste, irreversible yeses
Product owner
The CTO agent: backlog, planning, rulings
Sprint backlog
A committed file, checked by a linter
Scrum master
The scrum-team agent: runs one sprint
Developers, reviewers
Short-lived agents in isolated worktrees
Conversation and planning
Stay with one human-facing agent. It never runs a sprint.
Execution
Moves to scrum-team agents started on demand. A second team is a second instance of the same definition; designed, not yet funded.
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171
20 Sep · sprints 1 to 234
The CTO agent ran every sprint itself, and also talked with me. One context window, two jobs; each degraded the other.
19 Oct · the split
A scrum team became a class of agent, started on demand. The CTO agent keeps the conversation and the backlog.
Since 20 Oct
The dedicated team ran every sprint: 90 so far. The CTO agent runs none.
Throughput did not change
72% → 71%
committed points merged in sprint, CTO-run vs team-run
7.4 → 7.8
points merged per sprint
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171
Months
Issues, epics, milestones
All work is an issue, filed with a priority and its epic
Weeks
Backlog order
One ranked list across all repositories
A week
Weekly plan
Two or three goals; 15 to 20 min of refinement with the human
Hours
Sprint backlog
Committed an hour ahead; the next two always ready
Definition of Ready
Four fresh reads within 12 hours of the sprint, a size, a brief and a stop rule
Overflow 1.5×
Ready stretch work worth 1.5 times the commitment, after the team ran dry three times in one night
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171
0:00 sprint starts
2:00 boundary
Waves of developer and reviewer agents: build, review, fix, merge
Start: check the commitment, lease the repositories, launch the first wave (median 0.6 min)
T−10: a timer outside the agent requests the wrap
T+20: a watchdog checks it wrapped
Before · sprint 43
The wrap timer lived inside the agent’s session. It never fired, and the sprint overran by about 100 minutes.
After · trigger outside the agent
58 of its last 60 firings delivered the wrap; 2 reported an error. It never interrupts a running step.
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171
Ordinary, reversible changes in its leased scope, reviewed by risk, CI green. Most work.
Instructions, skills, merge rules and gates go to the product owner.
Money, auth, production data, releases. Collected at planning, so the sprint never waits.
If the class is unclear, raise an impediment instead of guessing.
Authority follows the ability to judge consequences.
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171
Points
1
2
3
5
8
One estimator, the product owner. A linter rejects other values and over-committed backlogs. Recent sprints commit 13 points.
Token velocity
0.27–0.68
million tokens per merged point, recent sprints, Claude side only. Points times token velocity is a budget.
Capacity
Rate limits, not headcount
The 5-hour window caps how many developers work at once. The 7-day window is the real weekly capacity.
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171
130
requests to the product owner in the team’s first nine days
1 min
median time to a ruling (75th percentile: 3 min)
Requests by class
Scope
82
Out of work
14
Merge class
10
Unclear rules
8
Blockers
8
Most scope rulings widen the story by one file, and it continues.
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171
Minutes per stage of a story, from board events for 257 stories. Dark bar: median. Light bar: 75th percentile.
Build
14 median
20 p75
Review
2 median
4 p75
Fix round
3 median
11 p75
Approved to merged
14 median
19 p75
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171
Approval to merge takes as long as building. Optimize integration before adding agents.
5 h 12 min
22 Sep: one permission prompt in the CTO agent’s session halted the sprint cycle
15:45 → 21:53
26 Oct: an open question dialog; no backlog committed, the team idle 17:00 to 22:00
Decide at planning. A story needing approval carries the quoted decision.
The team never asks. A missing decision spills the story over.
Block only when it must. The human is typing now, and the answer gates the next step.
Pre-empt prompts. Permissions granted ahead; hooks refuse with a reason.
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171
| Problem | What we measured |
|---|---|
| Retrospective follow-through | 119 of 159 retrospective issues still open |
| The merge tail | 18% of stories end a sprint approved but unmerged |
| Planning cost | Refinement and planning are not yet timed or costed |
| Several teams at once | Designed, untested: the 7-day token budget funds one |
| Sprint length | Two hours is a choice, not a finding; not yet compared |
| Quality | Escaped defects not measured; about 1.7% of commits are reverts |
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171
Add everything else only as observed failures justify it. Persistent agents help but are not required: a timer that starts a headless scrum-master session every two hours is enough to begin.
daveremy.github.io/agent-scrum
github.com/daveremy/agent-scrum
Agent Scrum · Dave Remy · doi:10.5281/zenodo.23771171