Skip to content

Research · November 8, 2026

When an assistant should speak, and when it should stay quiet

We read the proactive-agent literature to decide when an assistant should interrupt a person. Here is what the papers report, what we could not verify, and the rules we kept.

A proactive assistant has two jobs. It has to notice something worth saying, and it has to pick the moment. Most benchmarks score the first job. We care more about the second, because a correct suggestion at the wrong moment still costs the person attention. This note records what we read, how far we trust it, and what we decided.

What we read

The survey in docs/research/proactive/01-papers.md covers arXiv papers from 2023 up to the survey date. Its author states the limits: every entry was checked through the arXiv API, but only the abstract and metadata were read, never the full PDF. Anything that is not in an abstract is marked "not read". We keep that caveat here. When we say a paper "reports" a number, we mean its abstract does.

What the papers report

  • Proactive offers can be learned from real behaviour. The Proactive Agent paper builds a dataset from real human activity, has people accept or reject proposed tasks, and trains a judge from those labels. Its abstract reports an F1 of 66.47 percent for proactively offering help, above the other models it tested.
  • Finding the problem is the hard part. The PROBE benchmark splits proactive work into finding an unstated issue, naming the bottleneck and fixing it. Its abstract reports a best end-to-end score of 40 percent, reached by two frontier models.
  • Timing changes the cost. A programming-assistant study compared prompt-only help with a proactive agent. The abstract says proactive help raised efficiency but disrupted the workflow, and that presence indicators and interaction context reduced the disruption.

These are results from other groups' tasks, users and labels. They are order-of-magnitude evidence, not a forecast for our product.

What we could not verify

The survey states its own gap: it found no update of the classic mixed-initiative principles for assistants built on language models, and no dedicated studies of notification fatigue for them. Accuracy figures across these papers come from different labels, users and tasks, so we treat them as order-of-magnitude evidence only. We would rather leave a gap in our reasoning than fill it with a confident sentence.

What we decided

The decision code lives in packages/workspace-decide/src/interrupt/decide.ts. The comments in that file give the order in which it runs. This is design documentation of the code on the date of the research, not a product promise, and it can change.

  • Cost rules come first. A crisis-quiet marker comes first and silences only system-started items, never the person's own reminders. Quiet hours follow: only kinds the person marked urgent pass. Meetings, calls and Do Not Disturb allow no interrupting channel, and the card stays visible.
  • Fatigue is a rule, not a score. After two unopened dismissals within two hours, nothing interrupts.
  • Value is weighed last. Only an item that passes those rules has its value compared with the effective cost of interrupting, and the quietest channel that works is chosen.
  • Every outcome has a reason. The result is interrupt, hold or passive, with a reason code from a closed list in packages/workspace-contracts/src/interrupt.ts, so the person can be told why.

The order matters. A high-value message does not override a meeting.

What stays open

The papers measure other products, and our own acceptance data is still thin. The honest status is that the mechanism for when to speak is built and tested, and that the question of how much it helps a particular person is not yet measured. Until it is, the quiet setting is the safe one, and the person can choose it. For the approval side of the same idea, see the approval card entry and the proactive agent entry.

Where this comes from

  • docs/research/proactive/01-papers.md, dated 2026-11-07, arXiv abstracts only.
  • docs/research/proactive/06-dnd-model.md, the learned interruptibility model.
  • packages/workspace-decide/src/interrupt/decide.ts, the rule order quoted above.

Sources