We Are ETH Debugging AI Agents at Scale: Roman Engeler on Traces, Guardrails & Atla
Nov 6, 2025
In this discussion, Roman Engeler, co-founder of Atla and ETH Zurich alumnus, dives into the world of autonomous AI agents. He explains the importance of 'traces' for debugging AI decisions and shares insights on how his company enhances agent reliability. Roman compares startup cultures in London, Zurich, and Palo Alto, reflecting on his journey from research to entrepreneurship. He emphasizes the critical role of guardrails for safe AI deployments and offers advice for aspiring AI professionals while promoting ways for alumni to contribute back to ETH.
AI Snips
Chapters
Books
Transcript
Episode notes
From PhD Plans To Startup Speed Dating
- Roman moved from academia to startups after attending a startup speed-dating event and trying a technical product manager role.
- That brief experiment led to deep ownership at a fast-growing startup and kept him in the startup path.
Agents And Traces Defined
- An agent is a language model given autonomy and tools to plan and execute tasks across multiple steps.
- A trace is the step-by-step sequence of actions and tool uses that shows the agent's reasoning and decisions.
Scaling Supervision Challenges
- As agents become more autonomous and multi-step, supervising and understanding them grows harder at scale.
- Atla analyzes traces to surface patterns of failures, like tool misuse or missing confirmations, to prioritize fixes.
