Addy Osmani, the engineering lead at Google Chrome, has dropped a critical insight for anyone running AI agents in production: traditional code review is obsolete for agentic workflows. In his latest Substack post, "Agentic Code Quality," Osmani argues that because agents generate code at a volume no human can read, software quality now depends entirely on the constraints and harnesses surrounding the agent, not on manual inspection.
The End of Human-Readable Diffs
Osmani points out a fundamental scaling problem: "thereβs just too much code for anyone to read." When agents churn out hundreds of thousands or millions of changes daily, the bottleneck shifts from generation to verification. He notes that asking an agent to review its own work is unreliable; in one experiment, two reviews disagreed, and a re-run produced a third answer. You cannot gate a merge on a coin flip. Instead, deterministic constraintsβlike Sonarβs quality gatesβmust enforce a single bar for both humans and agents.
Constraints as the New Code Review
The core thesis is that "software quality now depends on the constraints you set around your agents." These aren't just unit tests. Osmani calls for a broader set of back-pressure mechanisms, including mutation testing, cyclomatic complexity limits, and architecture rules enforced by tools like ESLint. The goal is to create an environment where agents can fail without doing damage, providing trustworthy feedback loops rather than relying on a final, human-gated approval that slows down machine-speed delivery.
Scaling Through Verification Loops
When the volume of changes exceeds the capacity of verification tools, teams face three choices: scale the verification system, reduce the agent's output rate, or lower the quality bar. Osmani suggests a more nuanced approach: un-constrain where itβs safe and tighten constraints where it matters most. This allows teams to maximize throughput without sacrificing critical quality metrics like security or maintainability. The ultimate constraint remains human judgment, but it should be reserved for high-stakes architectural decisions, not routine diffs.
Key Takeaways
- Human code review cannot scale to agentic volumes; deterministic constraints must replace it.
- Agent self-review is non-deterministic and unreliable for merge gating.
- Quality is a collection of signals (security, performance, maintainability) enforced by back-pressure throughout the pipeline, not just at the end.
- Teams must deliberately choose where to apply strict constraints and where to allow agent autonomy to maximize velocity.
The Bottom Line
If your CI pipeline is still waiting for a human to read a diff generated by an LLM, you're already behind. Osmani is right: quality is in the harness, not the head.