Delegating Complex Workflows to AI Agents

The Three Interfaces for Agentic Work

AI agents can operate external software by utilizing three distinct connection methods. The Model Context Protocol (MCP) is the preferred, purpose-built standard for agent-to-app communication, though its availability depends on the specific service. APIs represent the traditional developer-centric approach, requiring credential management and documentation ingestion to allow the agent to understand the service's capabilities. Browser automation serves as the final fallback, where the agent uses a headless or visible browser session to interact with UI elements, forms, and buttons as a human user would.

Executing Large-Scale Projects

Delegating tasks involves identifying a tedious, click-heavy project and querying the agent on the best connection method.

• Email Marketing Overhaul: The author utilized a hybrid approach by combining an MCP integration for simple tasks like drafting broadcast emails with browser automation for complex visual workflow editing in Kit.
• Transcript Migration: The author used the ChatGPT desktop application in 'work mode' to navigate the Vimeo interface, download over 100 transcripts, and clean the text. The agent was then directed to a local folder to populate these transcripts into a Rails-based membership site.

Operational Strategy

Efficiency is measured by the reduction in human attention rather than raw execution speed. By creating custom 'skills'—reusable instructions or configurations for the agent—the author ensures that the agent documents UI quirks and maintains consistent formatting across repeated tasks. When using browser automation, the 'agent-browser' skill (developed by Vercel) is recommended over standard browser extensions for its superior ability to target DOM elements directly, which improves reliability and speed compared to visual-only clicking methods.