Agent trajectories as programs: fingerprinting and programming coding-agent behavior

Researchers have introduced new methods for comparing AI agents and understanding their behavior, enabling the identification of unique 'behavioral habits' and the development of procedural representations for agent problem-solving procedures.

The researchers, as reported in a paper submitted on June 15, 2026^[1], compared ten agents and found they could be identified by their behavioral habits with 85.7% accuracy when attributing unseen trajectories to the correct agent^[1]. They developed procedural representations for agent problem-solving procedures using an emergent vocabulary induction technique. The framework was applied to the SWE-Bench dataset to study the structural distinctness of agent trajectories. Additionally, the researchers introduced ProcGrep, a library for auditing and evaluating agents. Another study proposed a new approach to understanding AI model behavior by analyzing agent trajectories and identifying the 'intent-execution' gap between model assumptions and harness behavior^[2]. The researchers developed a simple and customizable harness called 'Simple Strands Agent' (SSA), which was able to reproduce or improve on the pass@1 performance reported by diverse model-provider families on popular agentic benchmarks. An analysis of 138k trajectories generated by SSA revealed model-level differences in problem-solving behavior^[2].