YC Startup School says coding agents still need human context in production incidents

ycombinator · x · 2026-07-29

A YC Startup School breakout on “software factories” shared two practical lessons about coding agents:

Models tend to suggest doing something even when the right fix is to step back and use higher-level context. In one production incident, a team pointed every coding agent at the problem and got a pile of confident but wrong suggestions until a principal engineer noticed the real fix was to block a single customer hammering an endpoint.
Anything that is hard for a lab to RL on is unlikely to improve as quickly as the rest of the stack. The speaker’s rule of thumb: if you add a workflow element that’s difficult to train against, don’t expect model performance there to keep pace.

The post frames coding agents as useful, but still dependent on human judgment, operational context, and workflow design.

More from coding & agent