Europe 2026
Why (Senior) Engineers Struggle to Build AI Agents — Philipp Schmid, Google DeepMind
Overview
Building AI agents presents unique challenges compared to traditional software development. Instead of acting as traffic controllers with predefined rules, engineers now function more like dispatchers, defining goals for agents without dictating every step. This shift requires embracing the non-deterministic nature of AI, treating errors as inputs, and moving from rigid unit tests to broader evaluations of agent reliability and success rates.
Who should watch
- AI Engineers
- Product Managers
- Builders working with AI agents
- Those struggling with agent reliability and development
- Engineers transitioning from traditional software to agent-based systems
Key takeaways
- Text and context, including multimodal inputs, are the new state, replacing traditional data structures and Boolean flags for representing information.
- Control must be handed over to the agent, moving away from predefined workflows to dynamic, goal-oriented interactions.
- Errors should be treated as inputs, similar to user input, to enable graceful recovery and avoid costly restarts of long-running agent processes.
- The focus must shift from unit tests to evaluations (evals) that measure the frequency of successful outcomes, acknowledging the non-deterministic nature of agents.
- APIs and tools need to be designed to be agent-ready, with semantic interfaces and self-documentation, as agents lack the historical context of human developers.
- Embrace the iterative nature of agent development, accepting that models and agents will evolve, and design for recovery and continuous evaluation.
- *Software is disposable; we are going to rebuild many of the same things with better models.*
- *We need to design for recovery. Models are not perfect. Agents are not perfect.*
Notable quotes
Philipp Schmid notes that when building agents, *we define the goal on what we want the agent to do, but we don't define the exact step the agent needs to take to achieve that goal.*
Schmid also states, *We need to move from unit test to evals. We need to test how often something works because agents are only successful if they are really reliable.*
Unofficial community note. Prefer the recording for nuance.