Anthropic pushes autonomous agents

Anthropic offered a glimpse of self-improving AI systems, with a researcher describing agents that improved performance across 10 benchmarks without degrading elsewhere. That detail is notable because one of the hardest problems in iterative model improvement is avoiding regressions: systems can optimize for one task while quietly becoming worse, less safe or less reliable on others. Separately, Anthropic introduced a framework intended to let AI agents control hardware in labs and factories. Put together, the two developments point toward agents that not only reason over software tasks but also operate equipment, raising the stakes for evaluation, permissions, audit logs and physical-world safety controls.

Sources from the briefing

• An Anthropic researcher just gave us a peek at self-improving AITechCrunch
• Anthropic’s new framework will let AI agents control hardware - ComputerworldComputerworld