Use case · LlamaIndex
Human-in-the-loop approvals for LlamaIndex agents
LlamaIndex agents and query engines can call tools that change the world. Gate the risky ones with a single gate.require() so a human approves before the tool runs, and every decision lands in a tamper-evident audit trail.
await gate.require("record.delete", { id });
// the agent tool waits for approval, else fails closedWhere it fits
- ›Gate a tool in a LlamaIndex agent or workflow
- ›Approve only high-risk actions by policy
- ›Prove who approved each tool call over data
Frequently asked questions
How do I add human approval to a LlamaIndex agent?+
Call gate.require() inside the tool the agent invokes. The agent step blocks on the human verdict, proceeds on approve, and fails closed on reject or timeout — with a tamper-evident record and no custom callback plumbing.
Does it work with LlamaIndex Workflows?+
Yes. Because Seal is a plain async call, you can place it inside a workflow step or an agent tool wherever an irreversible action happens.