Four legacy repositories, years of drift between them, and a product that needed to become one thing. The conventional plan was a multi-engineer quarter. I wanted to see if one engineer could run the whole consolidation with an AI agent doing the implementation and me reviewing every change before it merged. Most of the work that made it succeed happened before the agent wrote any code.
The setup was the strategy
Before execution I spent two weeks building the harness: standards files so the agent follows each repository's conventions instead of inventing its own; custom commands for the repetitive moves; and verification loops — test suites, build checks, review checklists — that define "correct" mechanically.
The decisions that mattered
Keep a human review gate. Every change went through the same review I'd give any engineer's pull request; I budgeted for that from the start. The agent's throughput is worthless if you can't trust what merges.
Put the time into scaffolding. You can't do this with one giant prompt; it only works because the infrastructure is there. When the agent has the standards, the commands, and the gates, the prompts stay short and the output stays predictable.
Tests decide when it's done. 580 tests had to pass, builds had to stay green, and nothing merged on model confidence. Same principle as our content pipelines at work: nothing merges without passing the gates.
How it turned out
Fifteen work days from planning to QA handoff, one engineer, four repositories consolidated into two — and QA found zero bugs, vulnerabilities, or security hotspots. This became the template the AI Engineering Champions program teaches: put the effort into the harness before writing prompts.
All of it — architecture, harness, supervision, and the merges. The complete methodology, including what the agent got wrong and how the gates caught it, is in the full write-up.