Agents That Self Improve While You Sleep - Fuad Ali, Arize AI
Summary
This workshop details the self-improvement loop for AI agents, a critical process for managing the non-deterministic failure modes of complex agents. The loop involves instrumenting agent traces, using an 'always-on' agent (Signal) to detect and cluster failures, converting these failures into living regression datasets, designing robust evaluations (Evals), and running experiments to prove fixes. The goal is to automate the feedback cycle, allowing humans to focus on high-leverage tasks like eval design and validating critical fixes.
Key takeaways
-
The Self-Improvement Loop
12:19
The core workflow is Trace $\rightarrow$ Signal $\rightarrow$ Eval $\rightarrow$ Experiment $\rightarrow$ Fix. This process ensures that every failure detected in production is captured, analyzed, and used to improve the agent's reliability.
-
Signal: The Always-On Agent
20:15
Signal reads live agent traces on a schedule (e.g., every 6 hours), clusters recurring issues, proposes fixes, and automatically creates evaluation checks and data sets for continuous monitoring.
-
Living Fixtures and Regression Testing
30:31
Failures are captured and added to 'living fixtures' (data sets), which serve as dynamic regression tests. This ensures that fixes prevent recurrence of past failure modes.
-
High-Leverage Human Intervention
40:48
Humans should focus on eval design (imparting 'taste' of what is good), dataset curation (validating findings for fixtures), and reviewing proposed fixes (PRs), rather than manually reading thousands of traces.
Technical details
-
Agent Tracing and Instrumentation
1020s
Agents must be instrumented to send traces (e.g., using the AX CLI) to capture detailed information like start/end time, latency, tool calls, and model metadata. This is essential for understanding failure modes.
-
Evaluation Methods (Evals)
2032s
Three main types of evaluation are discussed: 1) Code checks (deterministic, for simple validation); 2) LLM as a judge (qualitative analysis using another LLM); and 3) Agent as a judge (most open-ended, leaving determination to an agent). Decision models offer a cheaper alternative, providing type choices or confidence scores.
-
Experimentation and Validation
2701s
The experimentation stage involves re-running the agent against the 'living fixture' (failure cases) and scoring the results to prove the fix works and hasn't introduced regressions. This can be integrated into CI/CD processes.
-
Context and Data Flow
3000s
The self-improvement loop requires all context—traces, human judgment (annotations), eval scores, and failure evidence—to be held together in a single system to feed the next iteration of the agent.
Mentioned resources
- Arize AI
- Signal
- Wonder Toys
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.