Verifying Completion

Agents say "I'm done" when they think they're done — but they're often wrong. A verifier checks the actual browser state before accepting a terminate action. If verification fails, the rejection is fed back as an error and the agent keeps going.

UrlMatchesGate

The simplest verifier — checks that the current URL matches a regex:

import { UrlMatchesGate } from "@omxyz/lumen";

const agent = new Agent({
  model: "anthropic/claude-sonnet-4-6",
  browser: { type: "local" },
  instruction: "Complete the checkout flow",
  verifier: new UrlMatchesGate(/\/order\/\d+\/confirmation/),
});

The agent can only terminate when the URL contains /order/123/confirmation. Otherwise it gets: URL "..." does not match expected pattern.

ModelVerifier

A separate model inspects the screenshot to judge completion. Uses a cheap model (Haiku) so verification cost is minimal:

import { ModelVerifier, AnthropicAdapter } from "@omxyz/lumen";

const agent = new Agent({
  model: "anthropic/claude-sonnet-4-6",
  browser: { type: "local" },
  instruction: "Complete the checkout flow",
  verifier: new ModelVerifier(
    new AnthropicAdapter("claude-haiku-4-5-20251001"),
    "Complete the checkout flow",
  ),
});

The verifier asks: "Has the task been fully completed?" and expects YES or NO. After maxAttempts (default: 2) rejections, it hard-passes to prevent infinite gate loops.

// Custom max attempts
new ModelVerifier(adapter, "Complete the checkout flow", 3)

CustomGate

Full control — write any async predicate over the screenshot and URL:

import { CustomGate } from "@omxyz/lumen";

const agent = new Agent({
  model: "anthropic/claude-sonnet-4-6",
  browser: { type: "local" },
  instruction: "Submit the contact form",
  verifier: new CustomGate(
    async (screenshot, url) => url.includes("/success"),
    "Must reach the success page",  // optional failure message
  ),
});

Action verification

Separate from completion verification, ActionVerifier runs heuristic checks after every action using CDP state inspection. No API calls — zero cost:

const agent = new Agent({
  model: "anthropic/claude-sonnet-4-6",
  browser: { type: "local" },
  actionVerifier: true,
});

What it checks:

Action	Verification
click / doubleClick	Did the click target an interactive element? Did the URL change?
type	Is an input element currently focused?
goto	Does the current hostname match the target?

Failed checks produce hints fed back to the model (e.g., "Type action may have failed — no input element was focused. Try clicking the input field first."). The agent can self-correct on the next step.