Human-in-the-loop approval gates
Autonomous until it matters. Then a human decides — and the agent resumes.
Without memory
Full autonomy is a liability the moment an agent can spend money, touch production, or delete data. But bolting a human onto every step kills the point of automation — and ad-hoc “are you sure?” prompts don’t leave an audit trail.
With Rekall
Gate rules pause a run when it reaches something consequential. A human approves or rejects from the inbox — with the agent’s full context attached — and the run resumes exactly where it paused. Compliance-friendly autonomy with a paper trail.
How it works
Define the gates
Declare which actions require approval — destructive operations, spend thresholds, production changes — as rules, not per-agent code.
The run pauses itself
Hitting a gate checkpoints the run into an input-required state and notifies approvers via the inbox and webhooks.
Approve and resume
The approver sees what the agent intends and why (its decision context rides along). One click resumes the run with nothing lost.
See it in action
// The run pauses itself at the gateawait rekall.execution.pause(run.id, {reason: 'Approval required: DROP TABLE legacy_users',})// A human reviews it in the approval inbox — with the// agent's reasoning attached — then resumes the runawait rekall.execution.resume(run.id)