Add human approval gates to AI agents

Put a checkpoint in front of the actions you cannot take back — one API call, a named human, and a record of every decision.

The problem

Autonomous agents are useful precisely because they act without waiting. That is also the risk: the same loop that files a ticket can issue a refund, delete a record, or email a customer. Most teams want a middle path — let the agent run freely for routine work, but stop at the handful of steps where a mistake is expensive or irreversible. Building that checkpoint by hand means queues, notifications, a console, permissions and an audit log, all maintained per framework. It rarely gets built well, and it rarely gets built twice.

How Raposa solves it

Raposa is a human-in-the-loop approval layer that turns any step into a gate with one call. Where the agent is about to do something high-stakes, it calls request_human_approval() with a description of the action. Raposa pauses execution, routes the request to the right approver, and resumes only on an explicit approval; a timeout is never treated as consent. Because it is one HTTP call, the same gate works from any framework — LangChain, the OpenAI Agents SDK, a custom loop, an n8n workflow — and moves with you if you switch.

Gates are configurable without code changes to the agent itself. Require a single approver or N-of-M sign-off from a named group; add reminders, escalation to a second person, and an audited expiry. Approvers act from a signed email link, a Telegram card or a Slack message — opening the message never decides, only the button does — so they never need access to your runtime. Every gate decision is sealed in a hash-chained, exportable audit trail, and an HMAC-signed webhook tells your systems the outcome.

What it looks like

The result is a gate you configure once and reuse everywhere — routine actions stay fast, risky ones wait for a person.