It’s all about the Agent Experience

No. 8 Δ 26 February 2025 # AI and strategy @ Peter Vermaercke

Now is the time to watch the space around extending chat assistants with your products and services. With Model Context Protocol and Web Applets as notable examples, we’re seeing more and more movement going towards a shared standard for integrating with an AI assistant. It’s one small step for man, one giant leap for the Agent Experience.

Context is King

We know LLMs know a lot about a lot of things. But they certainly don’t know everything. They especially don’t know a lot of personal things, things that are important to you. In order to be really helpful, AI systems need to have the right context.

Apple Intelligence leverages your iPhone’s content to go beyond what ChatGPT can do. But that’s still pretty limited. Sure, our phones hold a lot of our valuables but it’s not everything. What about the help articles on your company’s internal blog, the catalogue of books for 4 year olds at your local library or the Korean products that are available at your local supermarket? Currently no AI assistant can access these easily. But this is about to change.

Anthropic, the company behind Claude, proposed a standardised way to offer a specific context: the Model Context Protocol or MCP. Matt Webb wrote an extensive blogpost on Extending AI chat with Model Context Protocol. It allows you to extend the capabilities of Claude with any specific context you desire. It’s currently picking up a lot of speed as it’s also one of the ways to extend the capabilities of Cursor, a very popular AI code editor.

A USB-C port for AI applications

MCP sounds (and frankly is) complicated, but Anthropic describes the project as “a USB-C port for AI applications”. Just like USB-C allows you to connect your devices to various peripherals, MCP provides a standardised way to connect AI models to different data sources and tools.

Schematic of Model Context Protocol → modelcontextprotocol.io

Context is not just limited to data sources, it also entails logic and tools. Those tools are particularly useful for agentic flows: processes where an AI agent takes the lead. Agents can reason (think in loops) and use tools where appropriate. But obviously we need a way to define these tools and let agents know when and where they can be used. MCP tries to fix this.

But we’re not there yet. As with any format war, there are multiple contenders. Web Applets is another interesting contender currently being backed by Mozilla. It uses a standard web stack which makes the learning curve a bit less steep. As promising as these initiatives are, it’s highly probable another standard will prevail. Which one becomes the standard isn’t too important yet, what’s important is to see that there’s a standard coming.

When we look at the AI Agent ecosystem, we can see this format war is mostly about the tool and data source elements.

The Agent Experience

The Web Applets’ intro page asks the right question: why make a computer talk to another computer via a clumsy web interface when they can just talk directly? Just like User Experience is all about creating a delightful, efficient and frictionless experience for humans, Agent Experience does the same for agents.

In a recent blogpost Mathias Biilmann, CEO and Co-founder of Netlify, wrote about AX: Agent Experience.

We need to start focusing on AX or “agent experience” — the holistic experience AI agents will have as the user of a product or platform.

So how can we make the agent experience efficient and frictionless? Many people just focus on the communication channel: APIs, data formats etc. The fact that a (mocked-up) video of two agents switching to Gibberlink for easier communication went viral is proving this point. The channel is the trivial, easy part. OpenAPIs and Gibberlink are not silver bullets for AX. It reminds me of UX discussions where people just focus on the UI. The interface is just the vessel, the experience is more about what you offer and how you bring it.

Not the how, but the what

That’s no different in AX. If we offer an agent the same data we offer humans but through an API, we’re not far off from just scraping webpages. We need to rather look at the strengths and weaknesses of an AI agent, and tailor what and how we deliver it to their needs.

Let’s look at some of the main differences:

Information processing: AI Agents can keep a much larger context window than humans. It makes sense to deliver more background information than we’d normally deliver to humans. Combine this with a much higher processing speed, AI agents can also handle more data faster, practically near-instant. This means agents can benefit from receiving much more data at once, compared to humans.
Learning and adaptation: AI Agents can learn from vast datasets, whereas humans learn from experience. But humans are much better at recovering from something unexpected and figuring out how to solve it. This means error handling for agents may need to be differentiated.
Cultural and social understanding: AI Agents can simulate social understanding based on learned data but do not possess intrinsic cultural context. This means agents may need to be guided more towards an appropriate next step in a certain situation compared to a human.

So instead of just focusing on the channel, focus on what you need to deliver to agents to optimise their experience: the right amount of the right content, in the right format at the right time.

Towards an App Store for Agents

Now that every day we release thousands of agents into this world, we also need a scalable way to let them know what tools and sources they could use. We need an index or tool store of some sorts, where agents or assistants can find relevant tools for their task. At the moment each vendor has something already or is building something. Doing this per vendor is not a great solution. We’ll need to work towards an open, safe and scalable index to provide agents with the right tools.

Lemni offers a store for actions that an agent could use. This is still targeted at humans, of course. → Lemni

Towards a better, more open, and connected digital world

Mathias also reminds us we shouldn’t use this new technology to just bring more noise into this world, just because we can.

Software designed for AI agents has the potential to deliver exponential value. As an industry we must collectively focus on building an open agent ecosystem and designing thoughtful AX to create a better, more open, and connected digital world.

Keeping the focus on that last part will be crucial but at the moment I don’t see it happen much. Knowing that regular GenAI tools are consuming enormous amounts of energy, the energy waste of unsupervised AI agents could totally eclipse that. As product people, let us not lose that focus and guard human-centric and planet-centric values.

AX design for your organisation

With so much movement in this space it’s high time to start thinking about the Agent Experience of your products and services. How would an agent ideally access your products and services? What do they need to perform optimally and efficiently? Which data would they require and in which form? The how we can figure out later, let’s first define the what.