Operating model

1,000 PRs a month at Mark My Words

What it takes to merge 1,000 PRs per developer per month, and what we need decided to start.

Presented by
Mark My Words
Date
November 2026

Today

We merge about 9 PRs per developer a month.

PRs merged
292
1 August to 4 November, in markmywords
PRs per developer per month
9
Across roughly 11 developers
PR-Check median, minutes
30
The check every PR waits on
PR-Check p90, minutes
134
One run in ten takes longer than this
  1. GitHub merged PRs and PR-Check.yaml run history.

Why it is slow

Every PR waits on a slow, heavy check.

  1. 01

    Slow to go green

    PR-Check takes 30 minutes at the median and 134 minutes at the p90.

  2. 02

    Often red

    7 of the last 27 PR-Check runs failed, so people rerun and wait again.

  3. 03

    Heavy on every PR

    Each PR builds a Docker image and deploys Helm previews to two regions, dev and au-dev.

The target

1,000

merged PRs per developer per month

About 45 per developer per working day, about 110 times today. At 11 developers that is about 500 merged PRs a day.

What others learned

Teams at this scale changed the road first.

  • Anthropic

    Went from 2,500 to 30,000 merges a day and dropped the strict merge queue.

  • poteto at Cursor

    Build trust before running work in parallel. Fix each mistake at the strongest layer.

  • Amp

    Pave the road. Isolate each agent instead of adding permission gates.

  • Linear

    Moved off styled-components in over 1,000 PRs, using codemods and agents.

Our rule

Trust comes before parallel work.

More agents only help once we trust the checks. So we fix the merge path and verification first, then add agents.

The merge path

A fast check gates the merge; the rest runs after.

Today

PR-Check, 30 minutes at the median

Docker build and Helm previews in two regions on every PR

A person clicks merge

Someone notices main is red

After Phase 2

One ci-ok job: lint, types and changed-file tests, under 5 minutes

Only with the preview label; the image builds once, after merge

Stamp approval plus a green check merges low-risk tiers

Slack alert and an automatic revert; no merge queue

Fixing agent mistakes

Fix each mistake at the strongest layer.

  1. 01

    Architecture

    Make the mistake impossible to write, for example by moving data access into one layer.

  2. 02

    Lint and CI

    A rule the linter or CI enforces on every PR, with no one watching.

  3. 03

    Review bot

    The AI reviewer flags it before a person reads the PR.

  4. 04

    Skill

    A skill teaches agents the right pattern for that kind of work.

  5. 05

    Doc

    The weakest layer. A rule left only as a PR comment is a bug.

Team structure

One pod paves the road; everyone else leads a stream.

  1. 01

    A paved-road pod of 2

    Owns CI, the merge path, Stamp, the verify command, agent environments and the throughput dashboard.

  2. 02

    Build leads: everyone else

    Each owns one product stream end to end. They set intent and say no; agents write the code.

  3. 03

    A pilot cohort

    Billy and Callum run the new workflow first, in weeks 1 and 2, then it rolls out to everyone.

Rituals

Three habits keep quality up as volume grows.

  1. 01

    Weekly taste session

    30 minutes a week comparing two options on real MMW screens.

  2. 02

    Weekly on-call

    One person each week owns main going red, so it never waits for a volunteer.

  3. 03

    Incidents become checks

    Every incident ends with a new check, so the same mistake cannot merge twice.

The plan

Four phases over about 12 weeks.

A one-day baseline comes first, so every phase is measured against today.

  1. Phase 1

    mmw-design repo

    One week. Brand, deck engine, slides site and UI registry, copied from amtec.

  2. Phase 2

    Fast merge path

    About 2 weeks. A ci-ok check under 5 minutes, previews off the hot path, auto-revert.

  3. Phase 3

    Guardrails for agents

    About 2 weeks. Boundary lint, a verify command with seeded logins, a new AGENTS.md.

  4. Phase 4

    Agent fleet and ramp

    Weeks 3 to 12. Pilot first, then cloud and scheduled agents, through gated steps.

Ramp gates

We ramp in gated steps, not all at once.

Step 1: PRs per developer a day
5
Weeks 3 to 4
Step 2: PRs per developer a day
20
Weeks 5 to 8
Step 3: PRs per developer a day
45
Weeks 9 to 12, about 1,000 a month
Revert rate to take the next step
<2%
And escaped bugs must not rise

How we measure

Four numbers, every week, per developer.

  • Merged PRs

    Counted per developer, every week, against the ramp gates.

  • Revert rate

    The share of merged PRs we had to undo. It must stay under 2%.

  • Time to merge

    How long a PR waits between being opened and being merged.

  • Escaped bugs

    Bugs that reach teachers, tracked in Done Bear each week.

Decisions

What we need decided

ActionOwnerBy
  1. 01

    Confirm markmywords is the target

    Not markmywords-new or a rebuild, so throughput rises this quarter

    Owner: Plan owner

    By: Phase 0

  2. 02

    Sign off the team structure

    It changes people's roles, so leadership agrees before the pilot

    Owner: MMW leadership

    By: Phase 0

  3. 03

    Set a monthly agent and CI budget

    Unknown today, and not unlimited

    Owner: MMW leadership

    By: Phase 0

  4. 04

    Confirm amtec code can be reused

    Code copied from amtec into an MMW-owned repo needs the go-ahead

    Owner: Plan owner

    By: Phase 0

  5. 05

    Choose the domains and Vercel team

    For example slides.markmywords.au and ui.markmywords.au

    Owner: MMW

    By: Phase 1