Mardu Swanepoel on AI Agent Best Practices

In a presentation at AI Engineer Europe, Mardu Swanepoel of Flinn AI shared insights into the core characteristics that define effective AI agents. Swanepoel articulated that the most successful agents share four key traits: focused modes, transparent execution, personalization, and reversibility.

The Pillars of High-Performing AI Agents

Swanepoel began by referencing a Pablo Picasso quote, "Good artists copy, great artists steal," to illustrate his point about learning from existing successes. He explained that "stealing" in this context means deeply understanding and integrating effective strategies into new creations, rather than simple imitation.

He then outlined the four key patterns observed in top-tier AI agents:

• Focused Modes: These agents operate within a "constrained scope for dedicated tasks." By narrowing the agent's focus, the output quality for specific functions is improved. This approach guides the user's expectations by clearly defining what the agent is designed to do. Examples shown included agents for coding (Cursor), knowledge workers (Claude Cowork), and legal professionals (Harvey).
• Transparent Execution: The principle here is to "make every step the agent takes visible." This transparency is crucial for building user trust, allowing them to understand the agent's reasoning and actions. Tools like Claude Cowork were highlighted for their ability to show progress and the context of their operations, enabling users to intervene if necessary.
• Personalization: This involves tailoring agent outputs to "user conventions, preferences, etc." Swanepoel emphasized that personalization significantly increases the "speed to understanding" for users. Agents that effectively incorporate user preferences and context are more likely to produce relevant and useful results, avoiding the pitfalls of generic outputs. Harvey's playbook system for legal firms was cited as an example of this.
• Reversibility: The final key trait is the ability for "every agent action to be easily undone." This feature binds the cost of mistakes, making users bolder and more willing to take risks with AI tools. By allowing users to easily revert actions, agents can provide a safer environment for experimentation and learning, ultimately leading to more efficient problem-solving. Swanepoel demonstrated this with Cursor's ability to undo changes at a granular level.

Swanepoel concluded by stressing that these four principles are fundamental to building AI agents that are not only functional but also trustworthy and user-centric. By "stealing" these best practices, developers can engineer more effective and reliable AI solutions.