Agents that
don’t wait
to be asked.

• Trigger 1

  Time

  The agent runs on a schedule it keeps for itself. Every fifteen minutes, every Monday at nine, every quiet hour past midnight.

  time.every("15 min", agent)

• Trigger 2

  Change

  The agent watches the world for a delta. A ticket moves, a record updates, a file appears — and it wakes the moment it happens.

  world.on("change", agent)

• Trigger 3

  Message

  Someone — a human, another agent, a system — addresses the agent directly. It answers in its own time, not on a polling cycle.

  inbox.on("message", agent)

A truly proactive agent listens for all three. Pick one and you’ve made a smarter cron job; pick two and you’ve made a chatbot that polls. The composition is what counts.

1. 01

   Wake-ups are infrastructure

   Polling is easy; push is hard. Stable URLs, signature schemes, normalized events, durable triggers — none of it ships in a model SDK. Someone has to build it.

2. 02

   State is harder than it looks

   Between wake-ups the agent has to remember what it saw, what it acted on, what it’s still in the middle of. Most agents wake up amnesiac and re-read the world from scratch.

3. 03

   Restraint is a research problem

   An agent that fires too often loses trust faster than one that misses things. Calibrated restraint is a known-hard problem even at the frontier — GPT-4o tops out around 65% on it.

• N°31

  June 19, 2026 · 8 min

  Open SWE and the architecture everyone built twice

• N°30

  June 19, 2026 · 9 min

  Open-Inspect and the Ramp Inspect blueprint

• N°29

  June 16, 2026 · 9 min

  Tasklet skips the workflow builder

01What is a proactive agent?

A proactive agent is an AI agent that acts without being prompted. Instead of waiting for a human to type a command, it wakes itself up when time passes, data changes, or a message arrives — and decides whether and how to act. The defining characteristic is how it wakes up, not what model or framework it uses.

02What is the difference between a reactive and proactive agent?

A reactive agent waits to be invoked — it receives a prompt, executes, and goes back to sleep. A proactive agent wakes itself up based on triggers: schedules (time), data mutations (change), or incoming messages. Reactive agents poll on intervals and forget between runs. Proactive agents receive push events and maintain persistent state across wake-ups.

03What are the three triggers that make an agent proactive?

The three triggers are:

1. 1.Time. The agent runs on a schedule or interval it keeps for itself.
2. 2.Change. The agent watches for data mutations via webhooks and acts the moment something moves.
3. 3.Message. Someone — a human, another agent, or a system — addresses the agent directly.

A truly proactive agent listens for all three. Using only one or two yields a smarter cron job or a chatbot that polls.

04How do you build a proactive agent?

Building a proactive agent requires three primitives:

1. 1.Wake-up mechanism. A clock, a listener for change events, and an inbox for messages.
2. 2.Persistent state. So the agent remembers what it saw and did between runs.
3. 3.Durability. Checkpointing to resume after failure, idempotency to prevent repeated actions, spend control, and scoped authentication.

Together these form the infrastructure layer that sits underneath the agent's logic.

05Why are most AI agents still reactive?

Three engineering problems keep agents reactive:

1. 1.Wake-ups are infrastructure. Push-based triggers require stable URLs, signature verification, and normalized events that don't ship in model SDKs.
2. 2.State is harder than it looks. Agents need persistent memory across runs, not just conversation context.
3. 3.Restraint is a research problem. Knowing when NOT to act is as important as knowing when to act, and calibrated restraint remains difficult even for frontier models.