AI agent infrastructure boom: memory, observability, human-in-the-loop become new battlegrounds

Memory infrastructure for AI agents. Memory is a long-term asset for agents, applicable to any personalized service.

eBPF-based observability for AI agents, no code changes. Observability is a blind spot in agent ops; non-invasive solutions are portable.

An L402 Lightning toll booth for AI scraping agents. Charging for agent access, applicable to content paywalls and API billing.

The agent infrastructure layer is densely emerging: memory, observability, human-in-the-loop, permission control, model routing, etc., indicating that agent applications are maturing and need supporting underlying services.

dsh-im and stop-that-shit gained attention in the open-source community; the former connects IM bots to DeepSeek Harness, the latter stops behavior creep in AI coding agents.

Vercel released fx, an open-source coding agent, and updated AI Gateway to support DeepSeek V4 Flash Vision; Google DeepMind partnered with game studios to prototype AI gameplay.