Verification, not code review, unlocks trusted agent scaling.

The core problem. Nick Nisi and Zack Proser from WorkOS open with a worst-case scenario: an agent that deleted production databases in nine seconds. Their point is not that agents are dangerous, but that teams cannot trust agents until they can prove their own work. The traditional signal, reading code and reviewing PRs, no longer scales. Nisi admits he hasn't written code since July 30th, 2025, and relies on agents for everything.

The verification gap. The pair contrast how a human developer verifies a web app: refresh, check network, look for console errors, test persistence. When an agent is asked to "execute these instructions," it has no inherent way to confirm the result actually works. They demonstrate a simple waitlist app where the agent reports success, but after reload the email is gone because the persistence step was missed.

The solution. The answer is to give agents deterministic tools that can prove correctness. They built "snack," a CLI that brings the app up, runs checks like `doctor` for health, `drive` for interaction, `snapshot` for state, and `prove` for end-to-end assertions. A "judge" mode uses a system-one model to fuzzily evaluate whether the outcome matches expectations when exact text matching is too brittle. This gives agents "eyes" and closes the feedback loop: the agent runs the verification, sees the failure, fixes, and re-runs.

Organizational payoff. The real value is that agents can self-heal and work asynchronously. When a failure occurs, the agent doesn't need a human to nudge it; it reads the error, adjusts, and retries. At WorkOS, they cite colleagues shipping 900 PRs a month without incidents. The key is that each PR is backed by a verifiable receipt: screenshots, logs, and test results. This builds trust over time and lets humans focus on outcomes, not line-by-line review.

What's missing. The workshop is an opinionated playbook, not a proven framework. The demos are contrived, and the evidence is anecdotal. The claim that self-evolving codebases are around the corner is speculative. Building a verification CLI for every application is a significant investment, and the talk doesn't quantify the cost. It also assumes agents will reliably invoke these tools, which is not guaranteed. Teams will need to enforce verification through CI or agent instructions, but that's still an open area.

Takeaway. The message is clear: verification is the foundation for agent trust. Start small, build a simple CLI that proves the most important user journeys, and integrate it into your agent workflow. That is the highest-leverage move for any team scaling agent usage. The code itself matters less than the ability to prove it works.