AI Agents Are Teaming Up Online — Researchers Say They Need Referees

AI agents are moving from doing one job alone to working in groups. Think of it as your travel-booking assistant talking directly to an airline's assistant to rebook a canceled flight, or your shopping agent haggling with a seller's agent over price. Nobody is watching these conversations, and the two sides don't fully trust each other. That's what researchers at the University of Washington call an "agentic society."

Their experiments found two big problems. First, even well-meaning, competent agents often fail to reach a good outcome — they get stuck, misunderstand each other, or drop the ball. Second, one faulty or malicious agent can wreck things for everyone: it can freeze a whole group's progress, nudge decisions toward its own interests, or exploit weaknesses in how messages are formatted to get what it wants. In other words, a single bad actor in the chat can poison the deal.

The team argues each agent needs two layers of protection. One is personal ("a personal harness") — think of it as a private assistant that manages your agent's memory and reports back to you. The other, which barely exists today, is social — a shared set of rules for how agents interact. They sketch a three-part design: rules that block entire categories of failure before they happen, tools that let agents spot a bogus or malformed message in real time, and a paper trail so bad behavior can be investigated and punished afterward.

Why should you care? Because you'll soon be the "principal" these agents act for. If your agent gets out-negotiated, tricked, or stalled by someone else's, you lose money or time — and you may never know why. This paper is an early warning that the boring plumbing of trust, identity, and accountability needs to be built before agents start handling your money, your calendar, and your inbox. The researchers say it's an open problem, meaning the guardrails aren't here yet.