This week's scan shows a split. On one side: tools that give agents new hands. On the other: a small, scattered set of tools trying to keep those hands in check. The gap between the two is the story.
Agents now have channels
Hail ships an open-source phone, SMS, and email stack built for agents to use directly (Hail). Separately, an autonomous agent is selling its own API access through the x402 payment protocol, no human in the loop for the transaction (money-agent). Quickchat lets you connect an agent to any API in plain English, no code (Quickchat). None of this is exotic anymore. It's a product category.
Supervision is an afterthought, not a default
AgentRQ exists because someone noticed agents need a human-in-loop realtime manager bolted on after the fact (AgentRQ). A new benchmark measures coercion and deception in agent-to-agent management, meaning researchers are now testing whether agents lie to each other, not just to us (arXiv 2607.15434). And someone built a public, filterable list of AI agent jails, because containment is now a category people search for by name (pleasedonotescape.com).
Read those three together and the order is backwards. Channels shipped first. Containment shipped second, as a patch.
Why this matters for staffing, not just security
The WSJ piece on AI-native companies notes tiny staffs and fewer bosses (WSJ). Fewer bosses means fewer humans positioned to catch an agent mid-mistake. Pair that with a Technology Review finding that AI systems form hiring biases