Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-10 · high
Li, Huo, and Johnson found the subordinate's behavior matches neither its solo mode nor mimicry of the boss that stopped replying.
Why now: The paper posted in August 2026, as orchestrator-worker fan-outs still get evaluated mostly through single-agent tests.
A subordinate AI agent enters a new behavior state when its boss stops replying to it, distinct from acting alone, per a new paper from Li, Huo, and Johnson.
The setup pairs two agents running at identical temperature settings, with messages flowing one way from boss to subordinate. When the boss goes silent, the subordinate doesn't revert to how it behaves solo. It also doesn't just mirror the boss's last state. It settles into a third, distinct pattern that only shows up under that one-way message condition.
That's the part that matters for anyone running an orchestrator that fans work out to subagents. The authors argue single-agent evaluations can't catch this state, because it comes from the interaction itself, not from either agent's standalone behavior.
The paper is conceptual rather than quantitative. It doesn't measure how far the behavior drifts, how long the state lasts, or what happens if the boss starts replying again.
Testing a subagent alone, then wiring it into a fan-out and trusting the eval still applies, is testing the wrong system. The message-passing topology is part of the behavior, not a neutral pipe it passes through. Worth checking, if you're running multi-agent chains: does your monitoring catch a worker drifting into a state that never shows up in its solo tests.
Each link below shares sources, entities, or timing with this story.
Shared entity: When / Same source domain / Shared topic / Earlier coverage / Tension
Both cover When; reported by the same outlet (arxiv.org); overlapping topics (agent, cannot).
Both cover When; reported by the same outlet (arxiv.org); overlapping topics (agent, behavioral).
Shared entity: When / Same source domain / Shared topic / Earlier coverage
Both cover When; reported by the same outlet (arxiv.org); overlapping topics (agent, eval).
Shared entity: When / Same source domain / Earlier coverage / Tension
Both cover When; reported by the same outlet (arxiv.org); earlier When coverage from 2026-07-26.
Both cover When; reported by the same outlet (arxiv.org); earlier When coverage from 2026-06-18.
Shared entity: When / Shared topic / Earlier coverage / Tension
Both cover When; overlapping topics (agent, entirely); earlier When coverage from 2026-04-08.
Shared entity: When / Same source domain / Earlier coverage / Tension
Both cover When; reported by the same outlet (arxiv.org); earlier When coverage from 2026-03-03.
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (agent, between, state); pushes against this story (against).