Harness over prompts: Google's Ryan Lopopolo on AI coding

The post draws a line between two ways of getting more out of coding agents: coaxing the model with better prompts, or changing the environment it works in. Lopopolo, who says he has not opened a code editor since May 2025 and now works through natural language and judges pull requests and documents against organizational standards, argues the second path is the durable one. He recommends "shifting left" — moving intervention earlier, to repository documentation, linters, tests and evaluations — so standards live in the environment rather than in each request.

That logic also explains Google's advice against over-building a custom harness. Standard harnesses already ship primitives such as file reading, grep search and command execution, so effort is better spent on tool and context quality, which stays useful as new models arrive. Lopopolo adds that a task a model failed six months ago may now be within reach, making fixed model-specific procedures a risky bet.

The demonstrations put that into practice. Jacobson's closed-loop harness feeds failing test logs and stack traces back into the agent's working memory and repeats until tests pass, capped at 5 iterations to avoid infinite loops and rising execution costs. Sumita Kauran of Google argues heavy reasoning models are not always required, pointing to a balance across model, harness and knowledge layers — with the Google Skills repository, holding over 100 domain packages, as an example of injecting accurate context on demand.