AI digest: Agent infrastructure gets serious
ByteDance drops a proper agent orchestrator, Anthropic squares off with the Pentagon, and everyone's suddenly worried about AI security testing.
The week that AI agents stopped being demos and started looking like actual enterprise software.
ByteDance releases DeerFlow 2.0 SuperAgent framework
ByteDance just open-sourced DeerFlow 2.0, a “SuperAgent” harness that orchestrates sub-agents, memory systems, and sandboxes for complex task execution. This isn’t another chatbot wrapper - it’s proper infrastructure for multi-agent workflows. Finally, someone’s building the boring plumbing that makes agent systems actually work in production.
Anthropic sues the Pentagon over safety guardrails
Anthropic is taking on 17 federal agencies after the Defence Department labelled them a supply-chain risk for refusing to drop AI safety measures. The 48-page complaint reveals Claude is already embedded in classified Pentagon systems, but Anthropic won’t budge on their safety stance. This could set major precedent for how much control governments can exert over AI safety decisions.
OpenAI acquires Promptfoo for agent security testing
OpenAI is buying Promptfoo to bake AI security testing directly into their enterprise platform. The move signals that prompt injection, jailbreaking, and data leaks are real enough problems that frontier labs need dedicated tooling. Smart timing as enterprises actually start deploying these systems beyond pilots.
Andrew Ng’s team launches Context Hub for coding agents
Context Hub solves a proper pain point - keeping coding agents updated with fresh API documentation instead of stale training data. It’s the kind of unglamorous but essential infrastructure that makes the difference between agent demos and agent deployment.