A new evidence-graded maturity model for AI agents, AgentStripes, is proposed. It defines levels, process areas, gates, and evidence grades to assess an agent's reliability, safety, and observability.

A new benchmark measures the readiness of merchants to engage with agentic commerce, evaluating how well their websites allow agents to browse, decide, and prove purchases. The benchmark assesses the determinacy of terms like total cost, cancellation terms, and service dates, and reports on access for different types of browsers.

This paper explores the choice between open-weight and closed frontier language models for security work, finding that open-weight models can match the performance of closed models at a significantly lower cost. The authors evaluate 26 models on the CTI-REALM benchmark and introduce stage-level metrics to score agent outputs.

Humanize is a multi-agent orchestration workflow for agentic coding that uses judgement engineering to improve reliability. It involves human approval, builder agents, and reviewer agents to enforce mechanical gates and catch defects. Deployed on 118 public postmortems, it achieved 1,468 GitHub stars and top rankings in several contests.

Research on the impact of large language model turnover on screening for AI-assisted scientific writing, highlighting the need for re-verification of benchmark accuracy with every LLM release.

OnTrack is a real-time monitoring mechanism for LLM agents that compares agent steps and dependencies against recorded successful runs to alert users or block the agent. It can detect plan violations, loops, stalls, and repeated tool calls with varying levels of data access.