Use casesEngineering teams

One agent learns.
Every engineer ships faster.

Wrenfield builds fleet telematics software: a nine-year-old monorepo, 140 engineers, a coding agent on every machine. Ten engineers on the dispatch services ran memco for thirty days against a measured baseline. The platform team then rolled it out to everyone.

Listen: the Wrenfield deployment

2:19

Organisation
Fleet telematics software, Bristol, 140 engineers
Agents in use
Claude Code and Cursor on every machine, Copilot in review, Codex on CI fixes
Tools connected
GitHub, GitHub Actions, Linear, Sentry, Slack
Pilot
Ten engineers, the dispatch service cluster, thirty days
Memory networks
Company, platform, one per service cluster, personal
Residency
Managed tenant in the EU, single sign-on
Time to first proven value
Day six

01Starting point

Where Wrenfield started

Every Wrenfield engineer runs a coding agent on the monorepo. The agent writes the migration, fixes the flaky test, opens the pull request. It is good at the task and ignorant of the repo: nine years of conventions, a build system migrated twice, an integration suite that only runs in CI, a legacy HTTP client that silently retries writes. Each session learns some of this the hard way. Then the session ends and the next one starts cold.

The platform team measured the problem for a fortnight before buying anything. On the dispatch service cluster, 41 percent of agent-opened pull requests came back with a correction a reviewer had already written that month, and six comments accounted for most of them. A quarter of sessions spent their first ten minutes on the same dead path, trying to run the integration suite on a laptop. The CLAUDE.md file was 700 lines long and nobody could say which lines still held.

The bill was reviewer time, bounced pull requests and tokens, in that order. Two senior engineers spent their afternoons repeating repo conventions in comments. What bothered the CTO more was the trend. Spend on agents was rising every quarter and the agents were no better at Wrenfield than the day they were installed. Each model release lifted everyone for a month, and then the gap between what the agents knew and what the team knew was back where it had been.

02Day one

Day one

The pilot was shaped the way the field guide recommends: one team, one service cluster, one repeated workflow, one metric. Connection took an afternoon. Seeding took a day.

  • Claude Code and Cursormemco plugin on the ten pilot machines, signed in through the company identity provider. Ten seconds per engineer.
  • Copilot and CodexMCP connection in the shared agent configuration, so the hosted agents in review and on CI read the same networks as the local ones.
  • GitHub and ActionsReview comments, merge outcomes and CI results attach to insights as evidence. Provenance points at the pull request, not at a summary of it.
  • Linear and SentryTickets and incidents link to the insights that resolved them, so a lesson can be traced to the work that taught it.
  • SeedThe platform team read the last quarter of review comments and the CLAUDE.md into forty insights, each with a scope and a source. Lines nobody could vouch for were left out.
  • SparkNetworks, validation rules and the moderation queue set up in spark.memco.ai before the first session ran.
  1. CompanyEngineering standards, security rules, the approved library list and the release process.Every agent and every engineer.
  2. PlatformCI behaviour, the build system, infrastructure conventions and deployment runbooks.Every agent working in any repo.
  3. Dispatch cluster, one network per service clusterModule ownership, test fixtures, migration patterns, known dead ends and the conventions reviewers enforce.Engineers and agents with write access to those repos.
  4. PersonalHow each engineer wants branches named, commits split and pull requests described.That engineer's agents only.

fig. 01Scope map

Wrenfield memory networksNested scopes: the company network read by every agent, the platform network inside it, five service-cluster networks inside that with the dispatch cluster as the pilot, and personal networks that read everything above and are read by no one; the only route up is a promotion through a named approver.COMPANYevery agent, every engineerPLATFORMevery repoCLUSTER NETWORKS, ONE PER SERVICE CLUSTERDispatchBillingTelemetryFleetMobilePersonal · one per engineer · reads everything above, read by that engineer's agents onlyPROMOTIONNAMED APPROVERReads inherit downward · Writes stay local
Wrenfield memory networksNested scopes: the company network read by every agent, the platform network inside it, five service-cluster networks inside that with the dispatch cluster as the pilot, and personal networks that read everything above and are read by no one; the only route up is a promotion through a named approver.COMPANYevery agent, every engineerPLATFORMevery repoCLUSTER NETWORKS, ONE PER SERVICE CLUSTERDispatchBillingTelemetryFleetMobilePersonal · one per engineer · readseverything above, read by that engineer'sPROMOTIONNAMED APPROVERReads inherit downward · Writes stay local
Reads inherit downward. Writes stay in the network where the work happened. The dashed line is the only route up: a promotion, held for a named approver after the day-twelve lint insight.

03The first weeks

The first weeks

Week one was quiet. Reads added about six percent to token use on the pilot cluster and returned little, because the memory was mostly seed. The first reuse came on day six, and after that the lessons arrived from the corrections engineers were already making.

fig. 02Deployment timeline

Wrenfield: baseline fortnight to readoutTimeline from the baseline fortnight through connection, the first reuse on day six, the over-broad lint insight on day twelve, the task-runner change on day twenty-two and the readout on day thirty.BASELINE FORTNIGHTTHIRTY-DAY PILOTDAY 6First reuse, in CursorDAY -14Baseline measurementbeginsDAY 0Connected; fortyinsights seededDAY 4Locking migrationrejectedDAY 12Over-broad lintinsight; trust dipsDAY 15Pull request formatlessonDAY 22Task runner replaced;eleven insightsDAY 30Readout: roll out to140
Wrenfield: baseline fortnight to readoutTimeline from the baseline fortnight through connection, the first reuse on day six, the over-broad lint insight on day twelve, the task-runner change on day twenty-two and the readout on day thirty.Baseline fortnightThirty-day pilotDAY -14Baseline measurement beginsDAY 0Connected; forty insights seededDAY 4Locking migration rejectedDAY 6First reuse, in CursorDAY 12Over-broad lint insight; trust dipsDAY 15Pull request format lessonDAY 22Task runner replaced; eleven insightsretiredDAY 30Readout: roll out to 140
Spacing is proportional to elapsed days. The filled mark is the first reuse: a lesson one engineer's agent learned, applied by another's.
  • Day 4. A reviewer rejects an agent's migration that adds a non-nullable column to vehicle_positions, a table holding two billion rows.Schema changes on vehicle_positions ship as a nullable column plus a backfill job, never a locking migration. See PR 8812.Day six: a different engineer's agent, in Cursor, proposes the backfill pattern unprompted and cites the insight. First reuse. The reviewer has not written the comment since.
  • Day 5. An agent spends forty minutes and 1.1 million tokens trying to get the integration suite green on a laptop.The integration suite needs the docker-compose stack from the platform repo. Run unit tests locally and push for integration.Sessions touching integration tests stop taking the detour. Median time to a green run on those tickets falls by a third over the pilot.
  • Day 9. Sentry flags a regression: an agent used the legacy HTTP client, which silently retries writes.Use the client in lib/net. The legacy client retries non-idempotent requests. Incident 2231.The platform lead promotes it to the company network, so every cluster picks it up at once. The incident is the provenance.
  • Day 12. An insight promoted from the dispatch cluster to the platform network tells agents to skip the lint step. True for one generated package, wrong everywhere else.Skipping lint applies only to the generated protobuf package. Everywhere else, lint runs. Platform insights that touch CI wait for a named approver.Two pull requests fail CI and a reviewer says he has stopped reading the recommendations. The scope is narrowed, the platform network gets a validation rule, and the reviewer is asked to approve the next ten platform insights himself. Trust comes back slowly, which is the honest timeline.
  • Day 15. The same senior engineer corrects the pull request description format for the fifth time in a week.Pull requests in the dispatch cluster open with the Linear ticket, a one-line risk statement and the rollback step.First-pass acceptance on dispatch pull requests climbs from 59 to 78 percent by day thirty. The engineer gets the afternoons back.
  • Day 22. The platform team replaces the task runner across the monorepo.Builds use the new runner from the platform repo. The old make targets are gone.Eleven older build insights are retired the same day with the change as provenance. memco flags two more that contradict the new runner; both are rewritten rather than left to mislead.

04Governance

Governance

  • The platform team owns the rulesValidation rules for every network are set in Spark by the platform team. Product teams propose changes to their cluster networks; platform approves changes to the company and platform networks.
  • Review by networkCompany and platform insights wait for a named approver. Cluster networks serve an insight automatically once it has held across three reviewed pull requests. Anything touching CI, auth or payments is judged against a written rule and held, whichever network it lands in.
  • Scopes mirror GitHubA cluster network is readable only by the people and agents with access to those repos. Personal networks never leave the engineer. Nothing moves sideways between clusters without a promotion.
  • Provenance on every insightEach insight links to the pull request, review comment, ticket or incident it came from, with a track record of how often it has held and when it was last confirmed.
  • Audit, retirement and residencyEvery read, write, promotion and retirement is logged and exportable to the SIEM. Insights retire on evidence, not on a schedule. Memory lives in a managed EU tenant. The code never leaves GitHub.

fig. 03The write path

From a review comment to the next engineer's agentSwimlane showing a reviewer rejecting a locking migration, the agent proposing the lesson with the pull request as provenance, memco holding it until it has held across three reviewed pull requests, then serving it to the dispatch cluster where another engineer's agent applies it on day six.REVIEWERENGINEER'S AGENTmemcoANOTHER ENGINEER'SAGENTRejects a lockingmigrationday 4Proposes thelesson, PR 8812attachedVALIDATION RULEHeld until itholds on threereviewed PRsServed to thedispatch clusterProposes thebackfill patternunpromptedday 6Never writes thecomment again
From a review comment to the next engineer's agentSwimlane showing a reviewer rejecting a locking migration, the agent proposing the lesson with the pull request as provenance, memco holding it until it has held across three reviewed pull requests, then serving it to the dispatch cluster where another engineer's agent applies it on day six.REVIEWERRejects a locking migrationday 4ENGINEER'S AGENTProposes the lesson, PR 8812 attachedmemco · VALIDATION RULEHeld until it holds on three reviewed PRsmemcoServed to the dispatch clusterANOTHER ENGINEER'S AGENTProposes the backfill pattern unpromptedday 6REVIEWERNever writes the comment again
Solid fill marks the reuse that counts as activation. The dashed node is the validation rule: a cluster insight serves only after it has held across three reviewed pull requests.

05What compounded

What compounded

The thirty-day readout compared the pilot cluster with its own baseline fortnight and with the four clusters that were not connected. The decision was to roll out to all 140 engineers in month two.

MetricMovementWhy
First-pass acceptance59% to 78%Each review comment that became an insight is one comment no reviewer writes again. The unconnected clusters stayed within two points of their baseline.
Repeat-error rateSix corrections to oneFive of the six comments that dominated the baseline stopped appearing by week three. The sixth, about test naming, needed a second rewrite of the insight before it held.
Time to an accepted resultDown 31%On tickets touching tests or builds. The detours that cost the first engineer an afternoon are recorded with the reason they failed, so agents route around them.
Cross-person or agent learning gainDay six, then weeklyEvery reuse was by a different engineer or a different tool from the one that learned the lesson. Eight of the ten pilot engineers reused another's correction in the first month.
All-in cost per accepted resultTokens per PR down 38%Counted after the reads that memory adds. Week one cost more than the baseline. The crossover came in week two and the gap widened every week after.
Improvement with experienceStill falling, month threeA model release in month two lifted every cluster. The connected clusters kept their lead over the unconnected ones through it, because the memory of the repo stayed with the company.

fig. 04Connected and unconnected

First-pass acceptance, baseline to day thirtySlopegraph: the pilot cluster rises from 59 to 78 percent of agent-opened pull requests accepted without a known correction, while the four unconnected clusters move from 60 to 61.BASELINEDAY 3060%59%78%61%Dispatch cluster, connectedFour unconnected clusters
Agent-opened pull requests accepted without a known correction. Baseline fortnight and day thirty.

fig. 05Costs before it pays

Tokens per merged pull request, week by weekLine chart indexed to the baseline fortnight: week one at 106 as reads add tokens before they return anything, the crossover in week two, 62 by week four and 55 by month three.6080100BASELINECROSSOVER10655WK 1WK 2WK 3WK 4MO 2MO 3Tokens per merged pull request, pilot cluster
Index, baseline fortnight = 100, counted after the reads that memory adds. Week one costs more than the baseline; the crossover is in week two.

06What it cost

What it cost

  • Setup timeA fortnight of baseline measurement before anything was connected, which the platform team now counts as the most useful part. One afternoon for the plugin rollout and MCP configuration. One day to seed forty insights from the last quarter of reviews.
  • Who ran itTwo platform engineers, part-time. About two hours a week approving platform insights in month one, under an hour by month three. No new headcount and no memory curator.
  • Licence and tokensThe Team plan for the ten pilot seats, then an enterprise agreement for 140 seats with single sign-on and audit export. Reads added six percent to token use in week one. By day thirty the net bill per merged pull request was down by more than a third.
  • What it replacedThe 700-line CLAUDE.md, the pinned Slack thread of repo gotchas, and two senior engineers' review afternoons.

See what your coding agents learn in thirty days.

Join the waitlist

More deployments.

All use cases