Your agent writes code faster than you can review it, and its tests come back green without proving much. That is the problem I work on with engineers and tech leads.

I build and run my own company's software with Claude Code every day, under a written, test-first protocol. What "done" means is written as yes/no checks before the agent starts, every test has to fail once before it is allowed to pass, and the work is reviewed in a fresh context, never by the session that wrote it.

Before this I spent over six years at PTC on the Vuforia Engine (computer vision and AR), where a large part of my work was verification: the SDK test suite, benchmarks and benchmarking research. Before that I did a PhD in computer vision at UCL. In October 2026 I spoke at TestIstanbul on "What If We Tested the Tester?".

How we work
In the first call we take one real pull request from your codebase, ideally one you did not feel good about. By the end of the first week you have:
- a definition of done for your next agent task, written as yes/no checks
- one test you watched fail before the fix, so a green suite means something again
- a review routine that tells you which lines a human has to read and which a check can cover

By the end of the first month you have that routine running on your real work and one number you track, such as rework on agent pull requests, review time or bugs that got through. From there we improve it month by month, and between calls you can send me a diff in chat.

If you manage the team rather than write the code, we work on the rollout instead: one shared rules file, one review gate, the number you report upward, and how to keep people on board when they use agents very differently.

If your code cannot leave the company, we work from a redacted diff or a small repository that reproduces the problem. Sessions in English, Spanish or Catalan.