← Field Notes
OCT 22 · Paper · via arXiv Agent-to-AgentConfidence Signaling

A new study finds more agents hurt quality in live chat

The human in the chat is not passive. Adding agents between the human and the answer made things worse. If your product puts a human in the loop at runtime, each extra agent is a cost, not a benefit.

Read the source ↗ attached to Agent-to-Agent →
emerging

watching → emerging → established

Seen across products but the shape is still settling. Adopt with an exit: expect the anatomy to shift.

18 documented sightings since Aug 2026 — every one traces to a dated source.

See the evidence →

Machine summary of the source

A pre-registered study tested whether adding more AI agents to a task improves results. In back-office work, where agents hand off to each other without a human in the loop, more agents did help. In live chat, where a human is part of every exchange, more agents made things worse.

The reason matters for product teams. A human in the conversation adds friction that extra agents cannot absorb. Each agent hand-off costs something, and in a real-time setting, the cost exceeds any gain in coverage or specialization. The study calls this a dual-control setting: the human and the system are both active at the same time, and they pull against each other.

If you are designing an agent product where a human talks to the system directly, the study says to start with one agent and add more only when you have evidence that the task, not the conversation, is the bottleneck. The default assumption that more agents means better results does not hold in user-facing products.

The summary above is generated; the note at the top is the editorial judgment. Primary source ↗