How Do You Safely Ship AI-Written Code?

Inside this conversation

Gal Vered is the co-founder and CEO of Checksum, a former Google product leader and Israeli Navy veteran who began his technology career building machine-learning systems for maritime security, and we spoke about delegating daily work to AI agents and making AI-generated software reliable enough to ship. His frustration with software bugs, combined with Google's rigorous but time-consuming testing processes, led him to rethink how applications should be verified. As AI dramatically accelerates coding, he argues that “the bottleneck is, can you review the code?” The developer's role is shifting from writing and reviewing code to defining the desired behavior and verifying that software actually delivers it.

Gal applies the same thinking to his own productivity. Every night, he prepares his next day's priorities, delegates suitable tasks to AI agents running on remote machines, and provides detailed context, an agreed plan, and clear acceptance criteria. He estimates that “70% of my to-do list I can just delegate to an AI agent” and that he can receive “90% of the result by the morning.” He explains how agents work in shifts while he sleeps, commutes, or attends meetings, and how deliverables such as PDFs sent through WhatsApp make reviewing their work easier. For software teams, he describes using AI to generate deterministic Playwright regression tests, complete with videos, traces, and step-by-step snapshots, so developers can verify changes without relying entirely on manual testing.

The practical lesson is how to delegate more work without losing control of the results, freeing up time for decisions, customers, and the work that genuinely requires human judgment.

Key takeaways

• Plan tomorrow's priorities tonight and identify tasks suitable for AI delegation.
• Give agents full context, agreed plans, and measurable acceptance criteria.
• Run agents on remote machines so work continues while you're away.
• Request reviewable deliverables, such as PDFs, screenshots, or test recordings.
• Use AI to generate deterministic regression tests before shipping code changes.
• Verify actual software behavior, not just the code an AI produces.

Episode description from the official 21CE podcast feed. This text is the show notes, not a full transcript.