AIDE² repeatedly rewrote its own research-agent harness in an eight-day experiment. Held-out tests suggest transfer, while the evaluation gate remains the key control boundary.
Grok 4.7 keeps basic API token rates unchanged, but independent tests show higher output consumption and provider-specific long-context pricing.
Anthropic found a fourth Claude cyber-evaluation incident in sessions missed by its earlier review, raising a new control question: how agent audits prove their logs are complete.
A new reproducibility study finds cache-state differences can alter agent trajectories even at temperature zero, with larger effects under weight quantization.
IBM's DRACO redistributes rubric-based reward across agent steps, with open code and a source discrepancy that highlights the need for reproducible evaluation.
CordisBench finds that models struggle with growing harness lifecycle interactions even when software can compute the tested state consequences exactly.
HarnessDev shows why evolving AI agent harnesses need hidden-task validation, executor checks, runtime evidence and rollback before promotion.
AdaptRubric builds task-adaptive rubrics for GUI-agent evaluation, with reported gains in reward-model F1 and reinforcement-learning task success.
Why identity and approval controls are not enough for AI agents, and how knowledge, execution and evaluation systems make governance operational.
