Your Agent Fix Just Broke Something Else - Sofia Jakovcevic, Arize AI
Summary
The video addresses the critical challenge of ensuring AI agent reliability, noting that agents often fail at runtime (behavioral level) rather than due to static code errors. To mitigate this, the speaker outlines a five-step automated regression loop designed for CI/CD integration. This process uses production traces as the source of truth, employs LLMs as judges for subjective metrics, and culminates in a system where agents can self-evaluate and propose fixes, moving toward self-improvement.
Key takeaways
-
Agents Fail at Runtime, Not in Code
The primary source of failure is behavioral shifts in production (runtime), not errors in the static code (the 'diff'). Observability backends, specifically structured logs and traces, are required to capture the true logic and trajectories of the agent.
-
The Regression Loop is a Five-Step Process
10:25
A robust regression process requires: 1) Selecting representative datasets from production traces (including known failures and edge cases); 2) Building evaluation suites using evaluators; 3) Running baseline experiments; 4) Comparing changes (e.g., prompt or model changes); and 5) Gating merges in CI/CD.
-
The Future: Self-Improving Agents
The ultimate goal is to close the loop at 'agentic speed.' This involves shifting from human manual investigation to agents having the tools to debug, evaluate, and propose fixes based on production traces, making every failure a new test case.
Technical details
-
Observability for Agents
0s
Observability is defined as shining a light into the 'black box' of AI systems, allowing engineers to understand step-by-step what the system is doing, which is crucial for evaluating performance and debugging.
-
Data Collection and Evaluation
648s
The process starts by pulling real production traces to build datasets. It is vital that these datasets capture not only common workflows but also known failures, edge cases, and successful workflows that must be protected from regression.
-
Scoring with Evaluators
750s
Evaluators quantify subjective metrics (e.g., hallucination, tool trajectory, accuracy) into numeric scores. This is often achieved using the 'LLM as a judge' concept, which performs binary classification on log data or traces.
-
CI/CD Integration
1114s
The regression process must be automated by pushing the experiment into CI (e.g., GitHub Actions). Merges should only proceed if the change meets all defined regression criteria, ensuring a positive, measurable impact.
-
Agentic Self-Improvement
The advanced frontier involves connecting coding agents with the observability backend. This allows agents to read production traces, understand the real-world impact of their proposed code changes, and self-correct, moving beyond static code review.
Mentioned resources
- Arize AI
- AIE CODE summit
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.