CORNERSTONE PLAYBOOK

An agent that can actually use your software — even the legacy stuff without an API.

An agent that uses a real browser — logging into your internal tools, running multi-step workflows, scraping data from systems that don't have APIs. Some of your most painful internal tasks live in apps that don't expose APIs — county records portals, insurance verifiers, supplier dashboards, regulator filings. A computer-use agent logs in, clicks, types, scrapes, and reports…

The problem, precisely

Roughly 30% of every services business's internal workflow lives in software that doesn't have a usable API. Examples: pulling county recorder filings to verify ownership, checking insurance eligibility on a state Medicaid portal, downloading supplier inventory from a legacy distributor dashboard, submitting state-licensing renewals through a 1990s-era government portal, running reports from a 12-year-old ERP that exports only to printer. The standard responses are all bad. Hiring offshore VAs is slow, error-prone, and bad for audit defensibility. Building bespoke scrapers with Puppeteer breaks every time the target site changes its DOM.

How we build it

We build the agent on Anthropic's computer-use API running on Browserbase (managed headless browser infrastructure with session recording and proxy rotation built in) or Hyperbrowser as backup. The reasoning loop is Claude 3.5 Sonnet for the planner + executor; we add an OCR fallback for sites that render content as images (PDFs of scanned forms, captcha-adjacent friction). Workflows are defined as natural-language SOPs ('log into the county recorder portal, search for parcels owned by [name], download the latest filing for each as PDF, push the PDFs to the Drive folder /research/[name]') and then translated by us into a deterministic-where-possible / agentic-where-necessary script. Determinism wins where it can — for steady-state portals, we cache successful action sequences and re-use them, falling back to vision-driven execution only when caching breaks.

What you get

Title research firm case: 4 paralegal-hours/day reclaimed on county recorder pulls = $1,400/week × 52 = $73K/year.

A worked example

A 6-person title research firm replaced 4 hours/day of paralegal time on county recorder pulls with an agent run. Net savings: $1,400/week. Setup cost: $9,500. Payback period: 7 weeks. Bonus: audit defensibility went from 'we have a sticky note' to 'we have a recorded browser session per query'.

Cost and timeline

Starts at $9,500 build. Ongoing: $1,499/mo + per-run pass-through ($0.15–$1.20/run). Ships in 21 days for the first workflow · the rest phase in after launch. Delivered as Browser Automation Agent, available across Global.

Is this allowed? Don't most sites' terms of service forbid automation?

Most sites' ToS forbid automated scraping of public data, but rarely forbid automation that performs the actions a legitimate logged-in user is authorized to perform. We only build agents that operate on systems where you have a legitimate user account and the right to perform the actions in question. We do not build scrapers of third-party-public data without authorization.

What happens when the target site changes its UI?

Computer-use agents using vision + reasoning adapt to small UI changes naturally (button moves, label renames, layout reflows) — the cost is a slight latency increase. For major redesigns we get a Slack alert ('workflow X failed validation on its scheduled run'), inspect, and re-deploy in typically <24 hours. For high-priority workflows we also keep a deterministic-action cache that retries first before falling back to fully agentic execution.

How do you handle MFA / captchas?

MFA is captured via app-specific authenticator codes (TOTP) that the agent has access to via Doppler. Captchas escalate to a human via Slack with a screenshot; the human solves it in Browserbase's live-takeover UI without restarting the session. For sites with persistent captcha walls we re-evaluate whether the workflow is suitable for automation — sometimes the answer is no, and we tell you.