Token cost is the obvious problem but I think the harder one is merge quality. Once you solve the cost side, you end up with agents that are cheap to run but still merge things that look right locally and break in prod.

What's helped us most is treating agent-generated PRs the same as junior-authored PRs: elevated scrutiny on file churn, coverage delta, and CODEOWNERS compliance. Agents tend to touch many files per commit (high change entropy) and skip writing tests — both are historically strong predictors of production incidents regardless of whether a human or agent wrote the code.