Working definition
What is proactive AI?
Proactive AI anticipates needs and decides whether, when, and how to help without waiting for a direct request. Sometimes the right choice is to stay quiet.
The three-question test
If the answer to the first is no and the other two are yes, the behavior is proactive.
- 1
Did anyone ask for this?
The action is not a reply to a request for this help.
- 2
Did the system judge the moment?
It decided whether and when to step in, rather than following a fixed trigger alone.
- 3
Could it have stayed quiet?
Not interrupting was a real option, and sometimes the right one.
The action ladder
Proactive does not mean acting more. These are the choices an agent weighs, and staying silent is often correct.
- Stay silentNothing is worth an interruption.
- ObserveKeep watching and gathering evidence.
- AskCheck before assuming.
- SuggestOffer a next step.
- PrepareHave the work ready for review.
- ActDo it within agreed boundaries.
How it differs
| Type | Who starts it | Who judges the moment | Example |
|---|---|---|---|
| Chatbot | You | You | A question-answering assistant |
| Scheduled or rule-based automation | A rule you set | The rule | Cron jobs, reminders |
| Autonomous agent | You set the goal | The agent, during the run | Coding agents working on a task |
| Clarification | You make a request first | The agent, on whether to ask | Adjacent: asks about a request already made |
| Proactive agent | The agent | The agent | Offers help nobody requested, or stays silent |
A schedule alone is not proactivity: a recurring job you defined does not show that the system discovered an unstated need. Products built on configured triggers still appear in the Atlas, marked by where their initiative comes from.
How the Atlas labels work
The main contribution is a decision to help, stay silent, or evaluate that decision, without a per-event request.
For example: PASK · HALOClosely related but not unprompted: clarifying a supplied request, timing an answer to a standing question, or governing authorization.
For example: AskStream · LexPilotA component a proactive system can build on, such as memory recall, that does not choose user-facing help by itself.
For example: The Price of Scale · ProV-MemLabels currently cover 89 recent and selected papers. Older entries are unlabelled until reviewed against their evidence notes.
Six questions the field is working on
A paper can contribute to several at once. Asking counts as proactive when the agent raises something on its own; clarifying a request you already made is adjacent.




