Automating the maintenance pipeline

The problem

Six enterprise SaaS products, most of them inherited through acquisition, all carrying defect backlogs. The documentation that came with them was scattered across wikis, dead Confluence spaces and people’s heads. Picking up a ticket on an unfamiliar product meant a day or two of archaeology before writing a line of code.

AI coding agents were the obvious lever, and out of the box they were useless here. A two-decade-old desktop CRM builds on Windows. The iPad client needs real hardware to install and debug on. An agent that cannot reach the platform cannot do the work.

What I did

The first problem was reach. I built harnesses that provision and drive Windows virtual machines so agents can build the desktop CRM and validate tests against it, and instrumented physical iPad hardware so they can install and debug iOS builds. That turns “agents cannot work here” into a solved problem.

Reach without context still produces guesswork, so the next piece was a central knowledge base covering every product in the portfolio. It consolidates the inherited mess into one structured source and is wired into the agents, so they answer from real product knowledge. I wrote the bulk migration tool that moves legacy documents in with AI-proposed structure, and the per-product context layer each agent loads before it touches a codebase.

On top of that sit serverless pipelines that take an incoming defect ticket, reproduce the failure, analyse crash dumps on remote machines, implement a fix, open a pull request, run the tests and review the result. Engineers handle what the agents cannot.

The review step needed the most care. An agent writing its own regression test will happily write one that passes against the broken code, which proves nothing. So I codified a red-green protocol: every regression test has to be observed failing before the fix is restored, enforced by a validation agent that rejects any test never seen to fail.

What changed

Onboarding onto an unfamiliar product got roughly 70 percent faster. My own merged changes went from 41 in the second half of 2024, to 131 in 2025, to 179 so far in 2026, while closing 580 of 591 assigned defects. The pipelines are used across the team now rather than just by me.