Run agent tasks: the hosted loop

import { Browserberg } from '@browserberg/sdk';

const bb = new Browserberg({
  apiKey: process.env.BROWSERBERG_API_KEY!,
  baseUrl: 'https://browserberg.com',
});

await using session = await bb.sessions.create({ ttlSeconds: 600 });

// runTask blocks until the run is terminal. The reviewer's verification is a
// SEPARATE model call against a re-scraped page: the agent never certifies
// its own success.
const run = await session.runTask({
  task: 'Open https://example.com and report what the page heading says.',
  maxSteps: 6,
});

console.log('status:', run.status);
console.log('answer:', run.answer);
console.log('reviewer:', run.verification?.thoughts ?? '(none)');

import os

from browserberg import Browserberg

bb = Browserberg(os.environ["BROWSERBERG_API_KEY"], base_url="https://browserberg.com")

with bb.sessions.create(ttl_seconds=600) as session:
    run = session.run_task(
        "Open https://example.com and report what the page heading says.",
        max_steps=6,
    )
    print("status:", run.status)
    print("answer:", run.answer)

status: completed
answer: The page heading says: "Example Domain".
reviewer: The page data now shows the heading 'Example Domain' on https://example.com, matching the agent's answer. Since the claim is of type `done` and the answer reports success, and the page evidence confirms it, the claim is valid.