Cloudflare moved its Monetization Gateway into closed beta on Sept. 30, enabling sellers to charge agents per request for APIs, MCP tools, websites, and datasets through HTTP 402/x402; Cloudflare cites four customer use cases already in production, including agent payment for API access. This adds a commercial payment/authorization layer to agent tool use, though access is limited to eligible U.S. beta participants. Source: https://blog.cloudflare.com/monetization-gateway-beta/
Respan introduced span-01 and span-01 Lite, reasoning classifiers that score agent traces against user-defined behaviors (e.g. prompt injection, tool misuse, or loops) in a forward pass rather than requiring a general chat model in the evaluation loop. The vendor reports span-01 at 0.843 overall F1 on its behavior benchmark and Lite as free; those comparisons and quality claims are vendor-reported, and independent validation was not found. Sources: https://www.respan.ai/blog/introducing-span-1 ; https://www.respan.ai/docs/documentation/span-01/concept
Newly surfaced research, updated Oct. 1: PrivMeSA combines a local clinical agent with remote specialist consultation, RL-based disclosure control, and lesson memory that reuses generalized guidance across cases. The authors report up to +15.8 percentage points in task accuracy while reducing direct personal-detail disclosure from 98.0% to 0.2% of cases on their MIMIC-IV-ED benchmark; early paper claims, not independent clinical validation. Its local/remote collaboration and memory design offer a concrete privacy-preserving pattern for agent systems handling sensitive data. Source: https://arxiv.org/abs/2609.38458