Goldberg: coding agents are co-workers, not compilers

In a May 2026 gist, Yoav Goldberg pushes back on a popular framing of coding agents: that reading their output is as unnecessary as reading a compiler's assembly. His objection is not the usual one about non-determinism, which he calls a bad objection that misses the mark. It is that compilers and agents do fundamentally different jobs.

"Compilers translate a complete specification to a runnable program, while a coding agent translates instructions and previously produced code to newly produced code," Goldberg writes. Each compiler run is isolated: input in, translation out, and if the program misbehaves you edit the input and run it again. Coding agents are stateful. Every invocation reads the code produced so far, and in a pure vibe-coding setup the user issues a sequence of behavior-change instructions applied on top of a hidden state they are not allowed to observe.

Goldberg's real target is what this implies for teams. Peer code review among human developers, he argues, is "not about function, its about structure" — a check that a colleague's code will be readable and easy to modify later. The same question applies to agent output, and he is skeptical current systems can answer it: "there is no indication nor any reason to believe that current day coding agents can assess what code will be easy or hard for themselves to maintain."

The closing move is organizational. Multiple people instructing agents against a shared, unread codebase creates a synchronization problem peer review of shared code used to solve. Goldberg suggests smaller human teams, perhaps a single person commanding a team of agents, closer to a product manager over trusted developers than an engineer running a build.

"A coding agent is not like a compiler. It is like a co-worker, or an employee," he writes. "If you don't want to read the code it produces, it might be ok. But don't do it because you also don't read your compiler's output."