AAgentify
human review|how it works|login

Human review for risky agent work

Agentify is the reference product for a simple pattern: agents can draft, browse, debug, and propose actions, but the risky step pauses for a TabWorker review before anything real happens.

1. Agent reaches a trust boundary

Agentify keeps the agent moving until an action touches money, accounts, production data, customer messaging, or another side effect.

2. TabWorker receives the review

The app submits a structured approval request through the public @tabworker/sdk surface, not a private integration path.

3. A reviewer returns proof

TabWorker captures the decision, corrections, notes, screenshots, URLs, and audit metadata needed to resume safely.

4. Agentify resumes only after approval

Polling or a signed webhook wakes the workflow back up with an approved, corrected, rejected, or escalated decision.

Reference SDK gate

TabWorker stays the reusable API and SDK. Agentify shows the outcome: an agent product can ask for human approval, wait for the decision, and resume from the same structured contract another builder would use.

import { TabWorkerClient } from "@tabworker/sdk";

const tabworker = new TabWorkerClient({
  apiKey: process.env.TABWORKER_API_KEY!,
});

const approval = await tabworker.approveAgentAction({
  queue_id: process.env.TABWORKER_AGENT_APPROVAL_QUEUE_ID!,
  action: "send_customer_refund",
  summary: "Agent wants to issue a $120 refund for invoice INV-123.",
  risk_level: "high",
  context: {
    customer_id: "cus_123",
    policy_excerpt: "Refunds above $100 require human approval.",
  },
  proposed_payload: {
    amount_cents: 12000,
    reason: "service outage credit",
  },
});

const result = await tabworker.waitForReviewResult(
  approval.review_request_id,
  { intervalMs: 5000, timeoutMs: 10 * 60_000 },
);

if (result.decision?.decision !== "approved") {
  throw new Error("Agent action was not approved.");
}

Where Agentify should pause

  • billing, refunds, credits, and account changes
  • customer-facing messages and support replies
  • browser findings that need source proof
  • code changes before deploy or public release
  • AI research, enrichment, and classification rows
  • eval labels where LLM-as-judge is not enough

Agentify sells the visible workflow. TabWorker sells the infrastructure. Start with the TabWorker SDK, read the review queue guide, or return to Agentify CLI.