# What the Best Agents Share

Mardu Swanepoel, Flinn AI | AI Engineer Europe 2026 | 10:21

Source: https://www.youtube.com/watch?v=7CrPrHgoEYk
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/what-the-best-agents-share
Published: 2026-05-26
Tags: design, guardrails, memory, observability

## TL;DR
- Focus modes constrain an agent's tools and inputs so engineers can improve quality and users know what to expect.
- Transparent execution turns agent use from delegation into collaboration by showing progress, tool calls, inputs, and outputs.
- Personalization and reversibility help agents do what users actually want while making experimentation safer.

## Summary
Mardu Swanepoel studies four agents, Cursor, Harvey, Manus, and Claude, to identify patterns that can improve agent design. Focus modes narrow the action space for a task such as planning or debugging, which gives engineers a smaller system to refine and gives users clearer expectations. Transparent execution exposes the agent's progress, context, tool calls, and results so users can build trust and intervene before wasted work continues. Personalization gives the agent a firm's methods, memories, skills, connectors, or other working principles, helping it reach understanding of the user's intent rather than producing a merely fast answer. Reversibility limits the cost of mistakes through undo controls and parallel outputs, which makes users more willing to try higher-value tasks. Swanepoel uses Cursor, Harvey, Claude, and Manus to show how these patterns appear in real products.

## Key ideas
### Focus modes improve agent output by narrowing the action space
[01:32](https://www.youtube.com/watch?v=7CrPrHgoEYk&t=92s)
Swanepoel defines focus modes as specific agent modes that constrain the available actions and inputs. An engineer can remove tools, refine the system prompt, and optimize evaluations for a smaller space before expanding the agent's scope. Modes also set clearer expectations for users who may not know how to get the best result from a general-purpose agent. Cursor provides examples through a drop-down interface. Planning mode asks questions and produces a plan without writing code. Debug mode follows a hypothesis-driven process, including starting a dedicated debug server and sending logs there.

### Transparent execution changes agent use from delegation to collaboration
[03:29](https://www.youtube.com/watch?v=7CrPrHgoEYk&t=209s)
Transparent execution makes the agent's actions, sources, assumptions, and uncertainty visible to the user. Swanepoel says this builds trust because users can inspect how a result was produced instead of seeing only the final answer. It also creates an earlier point for intervention. If an agent reads the wrong Notion documents at step two, the user can stop it and choose another approach before more work is wasted. Claude exposes a progress or task list, the next step, the context being used, connected skills, and the inputs and outputs of tool calls.

### Personalization helps the agent understand the intended way to do a task
[05:39](https://www.youtube.com/watch?v=7CrPrHgoEYk&t=339s)
Swanepoel describes personalization as giving an agent the systems, knowledge, principles, and patterns a person or organization would use themselves. The aim is to increase the speed to understanding, since a fast output is useless when it misses the user's preferred approach or unstated requirements. Harvey's playbooks capture a legal firm's methods for tasks such as contract review, so the agent can work in that firm's way. Harvey also creates memories from instructions over time. Claude offers related mechanisms through skills, connectors, and connected systems.

### Reversibility limits the downside of agent mistakes
[07:34](https://www.youtube.com/watch?v=7CrPrHgoEYk&t=454s)
Reversibility lets users undo actions made by an agent. Swanepoel argues that bounding the worst-case cost makes it easier to decide whether to let an agent attempt a task. Lower downside makes users more willing to take risks and use agents for higher-value work. Cursor supports several levels of rollback. Users can accept or reject changes line by line, accept changes at the file level, or return to an earlier conversation state and undo later messages and edits. Cursor also lets users run multiple outputs from the same input with different models, knowing that all but the preferred result can be discarded.

### Cursor combines focused modes with fine-grained experimentation
[02:57](https://www.youtube.com/watch?v=7CrPrHgoEYk&t=177s)
Cursor's modes connect a constrained workflow to a clear user contract. Planning mode does not write code, while debug mode uses a defined process for investigating possible problems. Its rollback controls then make experiments easier to run because users can review changes at line, file, or conversation level. Swanepoel presents this combination as a way to try different outputs while keeping the downside bounded. The user can compare parallel results and retain the useful one without accepting every change the agent made.

### Harvey applies firm-specific methods through playbooks and Word changes
[06:57](https://www.youtube.com/watch?v=7CrPrHgoEYk&t=417s)
Harvey's personalization comes from playbooks that encode how a legal firm approaches work, such as reviewing a contract. It also stores memories from ongoing instructions and interactions. For reversibility, Harvey is described as a Microsoft Word add-in that uses Word's native API. Users can view and manage changes in Word in the same way a reviewer or editor would work with document changes. This keeps the agent's edits inside a familiar review process rather than making the user accept an opaque document transformation.

## Notable quotes
- "The first one is what I call focus modes." (01:32)
- "The crux of what we're trying to achieve here is to shift from delegation to collaboration." (03:48)
- "Optimizing for speed to understanding in the sense of really understanding all of the nuances and implicit things from the user, how it would have approached it, is really critical for agent to do the right thing and not just something." (06:21)
- "We're binding the cost of our mistakes." (07:58)

## Tools & references mentioned
- Pablo Picasso
- Cursor
- Harvey
- Manus
- Claude
- Microsoft Word
- Notion

## Who should watch
- You are designing a general-purpose agent and need practical ways to narrow its behavior before adding more tools and tasks.
- Your users distrust autonomous actions or often discover mistakes too late. The talk gives interface patterns for showing progress and tool activity.
- You want people to try agents on more valuable work without exposing them to an unbounded cost when an action goes wrong.

## Related talks

- [How We Build Effective Agents](https://aietalks.com/talks/how-we-build-effective-agents) (Barry Zhang, Anthropic, 15:09)
- [Building Agent Interfaces: Lessons from Chrome DevTools (MCP) for Agents](https://aietalks.com/talks/building-agent-interfaces-lessons-from-chrome-devtools-mcp-for-agents) (Michael Hablich, Google, 22:38)
- [Designing Agents (The Floor Is the Frontier)](https://aietalks.com/talks/designing-agents-the-floor-is-the-frontier) (Ben Hylak, Raindrop, 19:46)
- [Agents Need More Than a Chat](https://aietalks.com/talks/agents-need-more-than-a-chat) (Jacob Lauritzen, Legora, 14:21)
- [Your agent is blindfolded](https://aietalks.com/talks/your-agent-is-blindfolded) (Johan Lajili, Poolside AI, 09:58)