Key Takeaways

  • Demetrios Brinkmann argues that vision-driven computer-use models remove the need for custom developer APIs by interacting directly with graphical user interfaces.
  • Modern agents can handle un-API-friendly municipal tasks, such as booking local German DMV appointments through standard web forms.
  • For high-stakes transactions like flight bookings, Brinkmann maintains a human-in-the-loop checkpoint, letting the agent find flights while clicking the payment button himself.
  • Complex multi-step campaigns can run across separate applications over several days by combining UI execution with asynchronous email monitoring.
  • Founders can automate restricted platform interactions using Brinkmann's Multi-Step Computer-Use Outreach Workflow.

The Brinkmann's Multi-Step Computer-Use Outreach Workflow

  • Step 1: Data Acquisition Trigger: Direct the computer-use agent to navigate the locked-down platform (such as LinkedIn) and request a complete export of account interaction data rather than attempting restricted API scraping.
  • Step 2: Asynchronous Inbox Monitoring: Instruct the agent to set an internal reminder and continuously monitor email for the delivery of the packaged data export over a multi-day waiting period.
  • Step 3: Relationship Filtering & Shortlisting: Once data arrives, parse and analyze message history to isolate authentic, multi-turn two-way conversations and generate a curated list of qualified contacts.
  • Step 4: Hyper-Personalized Messaging: Generate tailored outreach copy for each contact on the shortlist referencing prior conversational context.
  • Step 5: Dynamic Media Kit Generation: Spin up individual personalized landing pages containing ready-to-use social copy, suggested writing prompts, and media assets to streamline sharing.

When This Works (and When It Doesn't)

This workflow works when automating complex, asynchronous cross-application workflows across platforms that lack open or permissive developer APIs. Instead of building brittle integrations or paying for gated enterprise API access, you let an agent interact with the visual interface just as a human operator would.

It breaks down when platforms introduce aggressive bot detection, unexpected interface changes, or session timeouts during multi-day tasks. If a site updates its layout midway through an execution loop, visual coordinates and field selectors can fail. As Chris Benson points out, there is still an intrinsic trust factor with autonomous execution. Letting an agent handle open-loop tasks without sanity checks can quickly create errors in public-facing communications.

What to Do With This

Pick one locked-down platform where your startup needs data this week. Direct a computer-use agent to log into your account, request a data export, and route the download link to your shared inbox. While waiting for the file, configure your agent to filter the resulting interaction records for active conversations longer than three turns, draft contextual follow-ups, and hold the final send queue for your manual review.