Use casesA startup
How Loomwright's agents stopped forgetting.
Nine people shipping a scheduling product for field crews with coding agents all day, and a founder who reviewed nearly every pull request because she was the only one who knew the rules. She connected memco in an afternoon. By day three an agent had reused a lesson. By week six a new engineer was productive in days.
- People
- Nine, seven of them engineers
- Agents in use
- Claude Code and Cursor, daily
- Tools connected
- GitHub, Linear, Slack, one shared repo set
- Memory networks
- Loomwright, product, personal
- Review gate
- Founder approval on everything for a fortnight, then automatic except billing and sync
- Plan
- Team, nine seats at $50 a month
- Time to first proven value
- Day three
01Starting point
Where Loomwright started
Loomwright is nine people building scheduling software for field crews. Every engineer runs a coding agent all day. Tickets come from Linear, pull requests go to GitHub, and the agents write most first drafts. The founder reviewed nearly every pull request herself, because she was the only one who knew all the rules: money is integer pence, the mobile app merges offline edits per field, calendar changes ship behind a flag until the Tuesday release.
Each session began from zero. An agent learned on Tuesday that the billing module rejected floats, and on Wednesday a different agent in a different terminal put them back. The founder counted one month: the same review comment written eleven times, two engineers keeping private prompt files nobody else read, pull requests going through three rounds where one would do. The team was spending about $1,800 a month on agent tokens and a good share of it on rediscovery.
The cost was her time. Two hours a day correcting work she had already corrected. A new engineer the previous quarter had taken a fortnight to learn conventions that lived in nobody's head but hers, and the agents never learned them at all. She was the bottleneck on review and the only memory the agents had.
02Day one
Day one
One afternoon, by the founder, with no engineer pulled off a ticket. The hard part was not the plugin. It was writing down the six rules she had been enforcing by hand.
- Claude Code and Cursormemco plugin on seven machines, one team account. Each engineer signed in with their GitHub identity.
- GitHub and LinearPull request outcomes, review comments and tickets attach to insights as evidence.
- SeedThe two private prompt files and the founder's six standing rules became twenty insights with scope and source. The Notion conventions page was read, and most of it was found to be out of date and left out.
- ValidationFor the first fortnight every insight waited for the founder. After that, repo-scoped insights served automatically, and one written rule held anything touching billing or the offline sync for her.
- SparkThree networks drawn in ten minutes. The founder checked the review queue over coffee.
- LoomwrightThe six rules: money, sync, flags, regression tests, rollout notes, migrations.Everyone, every agent.
- ProductThe scheduling domain: how shifts, crews and calendars behave, what customers have complained about, which edge cases bite.Every engineer's agents.
- PersonalHow each engineer likes branches and commits.That engineer's agents only.
fig. 01Scope map
03The first weeks
The first weeks
The first week was the founder approving her own rules back to herself. The first lesson learned from live work arrived on day two and was reused on day three by an agent in another terminal.
fig. 02Deployment timeline
- Day 2. An agent puts a float in a price path and the founder rejects it, for the twelfth time that quarter.The billing module stores money as integer pence. Any float in a price path is rejected at review.Day three: a different engineer's agent in Cursor converts to pence unprompted and cites the insight. First reuse. The comment has not been written since.
- Day 5. An agent adds a required column to crew_shifts without a backfill and staging falls over.Migrations on crew_shifts run inside a transaction and need a backfill step before a column becomes required.Held for the founder by the migrations rule, approved, served. The next four crew_shifts migrations pass review first time.
- Day 8. An insight about the mobile sync says offline edits merge last-write-wins per record. It is wrong, and a crew in the field loses an hour of changes on a test build.The mobile app merges offline edits last-write-wins per field, never per record. Whole-record overwrites lose a crew's changes.The founder corrects the insight rather than deleting it, so the wrong version and the reason it was wrong stay in the record. Sync joins billing in the always-review rule.
- Day 11. A customer-reported bug is fixed before a regression test is written, and regresses a week later.Tickets tagged customer-reported get a regression test before the fix, not after.Linked to both tickets. Customer-reported regressions stop in month two.
- Day 16. A pull request touching two services ships with no rollout note and is rolled back on the Tuesday release.Pull requests that touch more than one service carry a short rollout note in the description, or they are sent back.The founder stops being the only person who asks. Reviewers who are not her start sending work back with the insight linked.
- Week 6. A new engineer joins.The six rules, the product network and forty lessons about the codebase, read from the first session.Productive in three days instead of a fortnight. Her agent proposes the pence conversion and the backfill step on her first two tickets without being told.
04Governance
Governance
- The founder owns the rules, for a fortnightEvery insight waited for her approval for two weeks. After that, repo-scoped insights served automatically once they had held on three reviewed pull requests.
- Billing and sync always waitOne written rule, judged against each insight: anything touching money or the offline sync is held for the founder, whichever network it lands in. Everything else flows.
- Provenance on every insightEach insight links to the pull request, comment or ticket it came from. When the sync insight was wrong, the record shows what was wrong and who fixed it.
- Shared cloud tenantNine people and a managed tenant. No residency requirement, no audit export, no identity provider beyond GitHub. The governance is the review rule and the log.
fig. 03The write path
05What compounded
What compounded
The founder kept the month's count she had made before connecting and made it again at day thirty and day ninety.
| Metric | Movement | Why |
|---|---|---|
| First-pass acceptance | 52% to 81% | Measured on agent-opened pull requests. By day ninety the number was 86 percent and the founder was reviewing a third of them. |
| Repeat-error rate | Eleven repeats to none | The six standing rules stopped recurring by week three. New repeats now get written down the second time, not the eleventh. |
| Time to an accepted result | Three rounds to one | The founder's review time fell from two hours a day to forty minutes. |
| Cross-person or agent learning gain | Productive in three days | A new engineer in week six. Her agent read forty lessons about the codebase in its first session. Nobody onboarded her. |
| Time to first proven value | Day three | The pence conversion, reused by a different engineer's agent in a different tool. |
| Improvement with experience | Through a model release | A new model arrived in month two and lifted the drafts. The rules carried over because they were never in a prompt file. |
fig. 04Two counts
fig. 05Costs before it pays
06What it cost
What it cost
- Setup timeOne afternoon, by the founder. Most of it writing down six rules she had been enforcing by hand.
- Who ran itThe founder for a fortnight of approvals over coffee, then a few minutes a week. Nobody else.
- Plan and tokensThe Team plan at $50 a member, $450 a month for nine. Reads added about five percent to token use in week one. By day thirty tokens per merged pull request were down a third against $1,800 a month of agent spend, and the founder's hour and a half a day was the larger saving.
- What it replacedTwo private prompt files, a conventions page in Notion that nobody updated, and the founder's second hour of review each day.