Use case · AutoGen

Human-in-the-loop approvals for AutoGen agents

AutoGen's multi-agent conversations can call real tools — Seal gates the dangerous ones. Wrap the tool in gate.require() and the agent pauses for a human verdict before it acts, with every decision recorded in a tamper-evident audit trail.

await gate.require("shell.exec", { command });
// the agent waits for approval — or fails closed

Where it fits

  • Require sign-off before an AutoGen agent executes code or shells out
  • Gate tool calls that touch production systems or customer data
  • Prove which human approved each autonomous action

Frequently asked questions

How do I add human approval to an AutoGen agent?+

Place gate.require() inside the registered tool function. When the AutoGen agent invokes the tool, execution blocks until a human approves, so risky operations like code execution or production writes need explicit sign-off.

Does Seal replace AutoGen's human_input_mode?+

It complements it. human_input_mode gates the conversation turn; Seal gates the specific side-effecting action with a policy, fail-closed default, and tamper-evident audit trail that AutoGen does not provide.