Why (Senior) Engineers Struggle to Build AI Agents — Philipp Schmid, Google DeepMind

TL;DR

• 1

  Agents change the job from control to delegation — Schmid says traditional software engineers acted like traffic controllers, but with agents you’re a dispatcher who sets the destination and lets the model choose the route, even when the path looks strange.

• 2

  Text is the new state layer — Instead of rigid booleans and flags, agents work on semantic context, like approving a deep research plan while adding “focus on the US market and ignore California” in one step.

• 3

  Errors should be fed back into the loop, not restart it — If a 5- or 15-minute agent run fails, starting over wastes compute and loses context, so errors need to be treated like inputs the model can recover from.

• 4

  Unit tests aren’t enough; reliability comes from evals — Because the same prompt can produce different outputs, the real question is not whether input A always yields C, but how often the agent succeeds well enough to trust in production.

• 5

  APIs built for humans often fail agents — A method like delete item may feel obvious to backend engineers, but agents only see function schemas and docstrings, so tools need explicit, self-documenting semantic interfaces.

• 6

  Build for replacement, not permanence — Schmid closes with the “bitter lesson” that agent software is disposable and will be rebuilt repeatedly as better models and agent patterns arrive.