Jared Zoneraich on Building Coding Agents That Verify Their Work

Cognition builder-in-residence Jared Zoneraich talks with Peter Yang about agent design, evaluations and cloud coding workflows. He demonstrates Devin, a product built by his employer. The episode is sponsored by Riverside.

Zoneraich distinguishes agents built for customers from internal agents that help a team work. He argues that teams should account for model improvements and not assume today's elaborate prompting workarounds will remain their main advantage. He presents this as a heuristic and says choices are case-specific.

On differentiation, he points to product quality, execution and customer support, including engineers working directly with customers. He says showing an agent solve a real customer problem can be more persuasive than explaining the technology in isolation.

The guests discuss how to ship an initial agent without spending indefinitely on a perfect architecture. Zoneraich says a first version can be simple, but also acknowledges that reaching a reliable final result takes longer. His advice does not mean skipping appropriate evaluation before production.

For a personal assistant project, he asks Devin to create sanity checks for data recall, tool use and response length. The agent compares old and new outputs and iterates. The example illustrates a feedback loop; its usefulness depends on whether the tests represent the desired behavior.

Zoneraich describes Cognition's focus on software engineering and verification, and contrasts synchronous work with asynchronous cloud tasks. His statements about Devin's strengths and customer adoption come from a person working at the company, rather than an independent comparison.

He argues that spending on agents should be judged against useful engineering outcomes, not token volume alone. He identifies migrations, test coverage and work in existing codebases as important customer use cases, while Yang notes that generated output still needs to be good.

In a demo, he shows a small Markdown preview project and starts an application test. The agent runs in a cloud environment with a browser and attempts to check the page. The demonstration shows an activity and interface, not proof that every defect was found.

He then asks a coordinating Devin to launch ten independent redesign sessions for the landing page. Each runs in its own virtual machine and branch. The coordinator can send follow-up instructions, such as asking each session for screenshots, and compare the results.

Zoneraich proposes large migrations as a more substantial use for this fan-out pattern: split a large task into focused pieces that can run and be tested separately. He also describes model routing and cost tradeoffs as product work in progress; parallel work adds coordination and compute costs.

He mentions shared knowledge, skills and saved environment setup, along with repository documentation, as ways to make subsequent work easier. These are product capabilities he describes at the time of the interview.

The conversation closes with predictions about coding agents, short-lived software and voice interfaces. Zoneraich imagines a much simpler desk, but acknowledges the unresolved question of how people will see and assess results. These are views about a possible future, not demonstrated replacements for current tools.