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

A new framework can identify distinct coding agents by their problem-solving habits rather than just their benchmark scores, according to research submitted to arXiv on 15 June 2026. The methods attribute unseen agent trajectories with 85.7% accuracy and introduce a library called ProcGrep for auditing agent behavior [1]. The work, authored by Hamidah Oderinwale, compares ten agents and finds that each leaves a procedural "fingerprint" — a set of behavioral habits that persist across different tasks and contexts [1]. A probe over these signatures correctly matches an unseen trajectory to its originating agent at 85.7% accuracy, controlling for task leakage [1]. The researchers developed an emergent vocabulary induction technique designed to compress surface-level variation while capturing the structural quirks of each model's approach [1]. Applied to the SWE-Bench software engineering dataset, the framework revealed that agent behavior clusters by release period and distillation lineage. A distilled student model and its teacher exhibited a Jensen-Shannon divergence of 0.25, roughly half the distance observed between unrelated model pairs [1]. This finding arrives as large language models proliferate rapidly; the list of publicly known LLMs has expanded to include systems from firms such as DeepSeek, which trained its V3 model for a reported US$6 million, far below the cost of comparable Western models [8]. The paper argues that as more models saturate traditional benchmarks, evaluating how an agent solves a problem — not just whether it succeeds — becomes essential [1]. This perspective echoes broader concerns about algorithmic transparency. Algorithmic bias, for instance, can arise from opaque design choices or unrepresentative training data, compounding social inequities in criminal justice, healthcare, and hiring [5]. The European Union's Artificial Intelligence Act, adopted in 2024, represents one regulatory attempt to address such systemic risks [5]. Alongside the fingerprinting technique, the authors released ProcGrep, a library that lets developers audit and evaluate agents by examining their procedural traces in a top-down fashion [1]. The tool is intended to support task-aware model routing, agent monitoring, and finer-grained cost analysis [1]. The submission, sized at 1,107 KB, was posted through arXivLabs, a platform that enables community collaborators to build experimental features on the arXiv website [1].