Interpreting Agent Behavior

As commercial agentic systems such as Claude Code and Codex see widespread real-world deployment in early 2026, their behaviors have grown increasingly complex. These systems autonomously plan, execute, and iterate on multi-step tasks, generating vast amounts of behavioral data. Yet current evaluation relies almost exclusively on automated benchmarks: pass/fail metrics that reveal whether an agent succeeded but not how it behaved, why it failed, or how humans made sense of its actions. Bridging this gap between outcome-based metrics and behavioral understanding is essential for building agents that are not only effective but also interpretable, debuggable, and trustworthy.

IAB aims to build a community that brings researchers across disciplines, including interpretability, evaluation, HCI, alignment, and computational social science. These researchers are actively developing analysis methods such as benchmarking, trace analysis, red-teaming, corpus analysis, error analysis, qualitative analysis, grounded theory, and think-aloud methods. IAB provides a venue to consolidate these efforts into shared vocabulary, open datasets, and reproducible methodology.