What the Best Agents Share — Mardu Swanepoel, Flinn AI

TL;DR

• 1

  Focus modes beat “do anything” interfaces — Swanepoel says agents like Cursor get better results by constraining the action space into modes like planning or debugging, which improves both evals for engineers and expectation-setting for users.

• 2

  Transparent execution turns delegation into collaboration — Showing progress lists, tool calls, inputs, outputs, assumptions, and uncertainty helps users trust the result and step in early when the agent is going off track.

• 3

  Personalization is really about speed to understanding — He argues many agents optimize for speed to output, but the real win is helping the system quickly absorb a user’s implicit preferences, principles, and workflow.

• 4

  Harvey’s playbooks are a concrete personalization pattern — In legal work, Harvey lets firms encode their own review methods into reusable playbooks so the agent behaves more like that specific firm, not a generic assistant.

• 5

  Reversibility changes the ROI math of using agents — If users can undo changes at the line, file, or conversation level, as in Cursor, they are much more willing to let agents attempt bolder, higher-value tasks.

• 6

  The unifying goal is bounded risk with better trust — Across Cursor, Claude, Manifold, and Harvey, the shared design pattern is making agents easier to steer, easier to inspect, and less costly when they make mistakes.