Key Takeaways

  • An agent that completes 99% of a workflow and drops the final step creates more user frustration than a tool that never started.
  • Pre-built API plugins should handle standard data retrieval and high-frequency actions because they run fast and conserve tokens.
  • Generalist computer use serves as a guaranteed completion mechanism when specialized software integrations and custom hooks fail.
  • Platform architectures require layered tool interfaces so first-party and third-party developers can build fallback paths into autonomous workflows.

The Last-Mile Agent Trap

Most agent demos fail in production for a simple psychological reason: partial completion feels like total failure. When an assistant searches flights, selects seats, fills out passport details, and then breaks on the confirmation button, the user has to audit every step anyway. They saved zero time and gained extra anxiety.

Nan Yu points out this exact tension. “There is a huge difference between getting all of the job done versus getting everything except for the last mile,” Yu explains. “And the last mile honestly feels sometimes worse than just like it's a non-starter.”

If software promises autonomous execution, 99% reliability is functionally equivalent to zero. Users do not want a copilot that requires constant supervision at the finish line. They want task completion they never have to double-check.

Platform Layering: APIs First, Computer Use as Fallback

To solve this last-mile failure, software platforms cannot rely entirely on rigid APIs or purely on visual mouse clicks. Custom APIs break whenever a third-party service updates its schema. Pure computer use, while flexible, remains slow and burns through context tokens rapidly.

Tara Seshan outlines the solution as a layered hierarchy. Primary actions run through structured integrations, while visual computer control waits in reserve.

“I think we should offer an array of different tools, like perhaps a plugin to be able to get all your meetings data and understand what you need to do next,” Seshan says. “But if that doesn't work and those systems don't interface well, then there's always computer use as a next layer or as a fallback.”

Seshan argues that platforms should expose these tools as structured hooks. Developers can route common requests through optimized data connectors first. If an unexpected pop-up, authentication wall, or unsupported field blocks the path, the agent drops down to raw OS interaction. It clicks the button, dismisses the modal, and finishes the task.

This hybrid approach balances speed and certainty. The user experiences an interface designed around hospitality rather than technical friction. Seshan references designer Charles Eames to frame the goal: the user should feel like a guest in your home, where you have anticipated what they need and handled the messy mechanics behind the scenes.

What to Do With This

Audit your primary user workflow and list the exact steps where an automated tool or agent drops the handoff back to the human. If your agent stalls because a third-party tool lacks an open API endpoint, build a computer use fallback script to bridge that single barrier this week. Do not ship an agent feature until it can resolve its own edge cases end-to-end.