World's Fair 2026
Your Agent Failed in Prod. Good Luck Reproducing It. - Tisha Chawla & Susheem Koul, Microsoft
Overview
This talk addresses the challenge of debugging AI agents in production, where failures can be difficult to reproduce due to the inherent non-determinism of LLMs. The core thesis is that instead of chasing bitwise determinism, which is often unattainable, developers should focus on replayability and observability. This involves capturing the full state of an agent's execution to enable debugging and testing, even when the exact same output cannot be guaranteed on subsequent runs.
Who should watch
- AI engineers and builders working with agents in production environments.
- Product managers and engineers concerned with the reliability and debuggability of AI-powered features.
- Anyone who has experienced difficult-to-reproduce bugs in AI agents.
- Developers looking for strategies to improve the testing and observability of agentic systems.
Key takeaways
- Attempting to achieve bitwise determinism from hosted LLM APIs is a losing battle due to factors like floating-point math, concurrency, and Mixture of Experts (MoE) routing.
- The goal should shift from making models deterministic to ensuring replayability, meaning the ability to reproduce and debug a specific past execution.
- Observability is key: capture the full input and output of each step within an agent's workflow, not just the prompt sent to the LLM.
- Tools like Chronicle can be used to annotate method boundaries, recording all inputs and outputs to create detailed traces of agent runs.
- These recorded traces can then be used for deterministic testing by stubbing out certain nodes and replaying others, allowing for focused testing of code changes.
- Guardrails can be implemented on tools to enforce credibility, and these guardrails can be tested using replayability traces.
- It is important to keep generation-time variation alive, as this randomness contributes to the agent's capabilities and creativity.
- *The right question is how do I debug and retest a run I cant reproduce because determinism was never the North Star. Debugging was.*
Notable quotes
The speakers suggest that chasing text outputs is a losing battle and that the focus should be on ensuring the system executes the exact same state transition.
*The right question is how do I debug and retest a run I cant reproduce because determinism was never the North Star. Debugging was.*
Unofficial community note. Prefer the recording for nuance.