Descriptions:
Philipp Schmid, who works on agents and the Gemini API at Google DeepMind, delivers a tight ten-minute conference talk explaining why even experienced software engineers consistently struggle when they first build AI agents — and what mental model shifts are required to do it well. The talk is structured around five concrete differences between traditional software development and agent development, drawn from patterns Schmid observes both internally at Google and across external teams using the Gemini API.
The five shifts Schmid identifies are: text is the new state (LLMs can handle semantic meaning that rigid data structures cannot, enabling dynamic personalization impossible with boolean flags); handing over control (agents need goal-level instructions, not step-level instructions, and their sometimes-strange paths to correct outcomes must be tolerated); errors are just inputs (agent flows must treat failures as continuable inputs rather than reasons to restart, especially when a single run already consumed fifteen minutes of compute); moving from unit tests to evals (agents are non-deterministic, so reliability must be measured statistically across many runs rather than asserted per-input); and the dispatcher vs. traffic controller distinction (engineers must shift from controlling exact execution paths to defining goals and letting the agent route itself).
The talk is particularly useful for senior engineers who bring strong software instincts that can actively work against them when applied to agent architectures. Schmid’s examples — a customer support agent that dynamically changes a user’s intent mid-conversation, a deep research agent that semantically approves partial plans — are concrete and immediately applicable.
📺 Source: AI Engineer · Published May 30, 2026
🏷️ Format: Deep Dive
