Listeners, not speakers: new study flips agent communication with interruptible generation
lileics · x · 2026-09-23
Researcher dqwang122 shared an 8-part thread proposing to tackle agent verbosity from the listener's side: a speaker sends messages in chunks, and the listener can keep listening, interrupt and respond, or ask for clarification—interrupting immediately halts generation, cutting both output and waiting time.
- Why not just prompt models to interrupt? They stop too early: in meeting scheduling, prompted listeners interrupted after the first chunk 87.6% of the time, and missing details caused extra rounds and higher total cost. Stopping early ≠ communicating efficiently.
- Evaluation: three settings—2-agent text Pictionary, 3-agent meeting scheduling, and 3-agent MMLU-Pro debate—covering games, coordination, and reasoning, using Llama 3.1, GPT-4o, and Gemini 2.0 Flash.
- Core question: "enough information" depends on the listener; one agent may understand after 16 words while another needs more.
More from coding & agent
- TesterArmy raises $1.2M pre-seed to build AI agents that test coding-agent-built apps — fernandorojo · 2026-09-23
- Two personal AIs tried to schedule coffee: agent interop needs a protocol — signulll · 2026-09-23
- Devin's iOS beta "Cog" hits TestFlight: code from your phone — msg · 2026-09-23
- Dev credits Alchemy for letting agents debug real deployments, landing major perf wins — samgoodwin89 · 2026-09-23
- Hooking Codex up to iMessage lets you interrogate months of chat history — gregmushen · 2026-09-23
- Sentry founder David Cramer finds Claude Code's new sidebar confusing vs Codex — zeeg · 2026-09-23