What the Best Agents Share — Mardu Swanepoel, Flinn AI

Mardu Swanepoel, founder of Flinn AI, presents a comparative analysis of four leading AI agents — Harvey, Cursor, Manus, and Claude — to identify the shared design patterns that make them effective. Drawing on the Picasso maxim of studying deeply and making it your own, Swanepoel argues that practitioners should extract these patterns and apply them when building their own agents. The talk identifies four distinct patterns: focus modes, transparent execution, personalization, and reversibility.

The first pattern, focus modes, involves constraining an agent's action and input space by placing it into a specific operating mode — such as planning, research, or debug mode. Swanepoel argues this yields two compounding benefits: engineers gain a smaller, well-defined action space they can actually optimize (refining system prompts, dropping irrelevant tools, tuning evals), while users receive clearer expectations about what the agent will and won't do. Cursor is cited as the primary example: its chat interface exposes a dropdown to switch between modes, where planning mode produces plans and questions (no code written), and debug mode applies a hypothesis-driven approach with a dedicated debug server and log streaming.

The second pattern is transparent execution — surfacing the agent's internal reasoning, tool calls, inputs, and outputs to the user in real time. The core goal is shifting the user relationship from pure delegation to active collaboration. Swanepoel notes that transparency builds trust in outputs (akin to a colleague showing their working rather than just delivering a result) and enables earlier intervention when the agent goes off-track, reducing wasted compute and effort. Claude is highlighted here: it displays a live to-do list of completed and pending steps alongside every tool call's inputs and outputs. Manus is cited as using a similar task-progress interface with visibility into what sources the agent examined and what conclusions it drew.

The third pattern is personalization — equipping the agent with the user's own methods, heuristics, and domain knowledge so it optimizes for speed to understanding rather than merely speed to output. Swanepoel distinguishes the two concepts sharply: an agent that produces fast outputs misaligned with the user's intent is useless, while one that deeply understands the user's implicit preferences and ways of working produces genuinely useful results. Harvey exemplifies this with its playbook feature: legal firms encode their specific contract review methodologies into playbooks, and the agent executes tasks the way the firm would. Harvey also builds persistent memory from prior interactions. Claude is noted for its skills and connector system that expands the agent's contextual knowledge base.

The fourth pattern is reversibility — the ability to undo agent actions at varying levels of granularity to bound the downside of mistakes. When users know the worst-case cost of an agent action is recoverable, Swanepoel argues they become bolder and more willing to delegate higher-value, higher-complexity tasks. Cursor enables rollback at the line level, the file level, and the full conversation state level. It also supports running multiple model outputs in parallel from the same input — effectively planning to discard all but one — so experimentation is low-risk. Harvey implements reversibility through a Microsoft Word add-in that uses the native Word API to present changes in a familiar track-changes interface, letting legal reviewers accept or reject edits exactly as they would in a traditional document workflow.

Swanepoel concludes by emphasizing that these four patterns — focus modes, transparent execution, personalization, and reversibility — are observable across the best production agents today and represent a practical framework for any team building agents to benchmark against and incorporate into their own designs.