AI Boost X
On this page
  1. The result
  2. Four sessions, one job each
  3. Tests that had to fail first
  4. What kept the quality up
  5. Caught before anyone saw it
  6. How I work

How I build the Agent Harness, 10,700 lines changed in hours while keeping the quality bar

Sep 2026

ai-tools claude-code agent-workflows case-study

The result

Files Changed85
Lines added10,712
Total Commits46 (42 by the agents)
Test runs on the merged code614, all passing
Planted defects caught by the tests28 of 28
From merge request to mergeabout 4 hours 15 minutes
42 commits, 09:49–16:15 · merge request → merge ≈ 4 h 15 m · 4 foundation commits the night before.
The day on a line night before 4 commits 42 commits 09:49 16:15 merge request opened → merged: ≈ 4 h 15 m

My harness conducts multiple agent sessions that built a production feature in hours with adversarial reviews, and run comprehensive tests, keep high bar quality of code changes.

That volume should worry anyone who has reviewed code. It held up because the harness I run assumes every model will sometimes be wrong, and checks each claim before it reaches a person or a system we don't own.

Where the 85 files went — 85 files in total.
Where the 85 files went Helm charts 23 Service source 17 Tests 14 Pipeline stage 10 Design docs 7 Schema and build — the rest 14

Four sessions, one job each

One human and four agent sessions, by level Human Level 1 Level 2 Level 3 Me — the only human decides: revert · push · post · approve · merge Upper level reviews text before it reaches the client team Conductor Fable 5.1 reviews and gates Runner 1 · Epic 1 Opus 5.5 · epic 1 done keeps the spec, watches Runner 2 · Epic 2 Opus 5.5 · epic 2 in progress builds code, tests, charts Push · merge request decides watches review file GO file

The conductor, on Fable 5.1, reviews and gates. Runner 2, on Opus 5.5, builds epic 2: the code, tests and charts. Runner 1, also on Opus 5.5, keeps epic 1's specification and watches the conductor's writes. An upper level reviews the text before it reaches the client team. I'm the only human in the loop, and I make every call that can't be undone: revert, push, post, approve, merge.

Tests that had to fail first

614 test runs by suite — 252 + 38 + 252 + 72 = 614. 28 of 28 planted defects caught.
614 test runs by suite Unit 252 Pipeline-stage unit 38 Integration — real PostgreSQL + Kafka 252 Inside the built image 72

Before the tests were trusted, Runner 2 planted 28 defects in scratch copies of the code, one at a time. The suite caught all 28. Integration tests ran against real PostgreSQL and Kafka, and the gateway's tests ran inside the built container over a real TLS stream.

What kept the quality up

  1. Nobody checks their own work. Whoever wrote something never reviews it.
  2. Reviews are files. Each names what it checked. Seven review rounds ran before the merge request opened.
  3. Re-read the source before every claim. Tickets, merge request and branch, pulled again in the same turn.
  4. Approvals are files, too. Nobody acts on an "approved" in a chat line.
  5. Every write to someone else's system is watched live. 33 requests over three days, none unwatched.
  6. Guards are hooks. Off-limits paths are refused by the tooling itself.
  7. Context loss is planned for. Each session keeps an anchor file. Three context resets that day; no step lost.
One watched write, step by step
  1. I write the command, the resource, the expected effect, a snapshot, and the rollback.
  2. Watcher reads them and records a GO.
  3. The command runs once.
  4. The result is read back from the server.
  5. It's compared by hash.
~2 minper write
33requests
30watched with a named watcher
3voided
0unwatched

Caught before anyone saw it

WhatCaught by
Two regressions introduced by an earlier round's fixesa review round
A safety-hook fix that passed only its own 51 tests57 attack cases written by Runner 1; the fourth build held
20 findings on the previous feature's specificationa standing red team; all 20 accepted

How I work

This is how I run AI agent teams on production work. Beyond this one day, I've implemented multiple agent-driven campaigns in production at scale. More on the method in AI Boost.

Discuss it on LinkedIn.