THE LOOP
From a stall to a reply the agent can use.
Checkpoint is a pager, not a handoff into a general chat. Each page has one question, one person, and one answer that goes back to the working agent.
01
The agent stalls
The agent is mid-task and cannot proceed without a fact, an approval, or a judgment. It stops instead of inventing the missing piece.
02
Checkpoint routes the page
The page names what is needed and why the agent stopped. Checkpoint sends it to the person who owns that answer — by role, ownership, or who is on call.
03
The human replies
That person sees the task context and answers the specific question. The reply is the context the agent was missing, not a new conversation to summarize later.
04
The agent continues
The answer returns to the same task. The agent proceeds with what the person said, instead of a guess it made while waiting.
Questions
- What is Checkpoint?
- Checkpoint is a human-in-the-loop pager for AI agents. An agent calls it when a task needs a fact, an approval, or a judgment that only a specific person has.
- When should an agent page a human?
- When continuing would mean inventing the missing piece. The agent stops, sends the question and the task context, and waits for the person who owns the answer.
- Who receives the page?
- The person matched by role, ownership, or on-call. The agent does not guess an email address.
- Is there a public API?
- Not yet. Checkpoint is in early access. The call an agent would make is described on the For agents page.
Agents do not pick an email address. See what a page asks for.