Executive Summary
ML6 went to the AI Engineer summit in Paris. We are confident that agents are here to stay, and MCP will be at the forefront of this trend. While MCP adoption is skyrocketing, its potential remains heavily underutilized. And while many agent projects are still explorative, those that find a suitable use case can radically transform their processes through iterative engineering.
Agents are still a really hot topic. Personally, we felt the agent fatigue before this event. There are dozens of agent SDKs, a handful of new models every month, and even more companies beginning to offer agent-focused services. After all these groundbreaking announcements, you can become numb to it. An excellent overview can be found in Swyx’s opening presentation.
However, after going to the AI engineer event, we are now more confident than ever that it’s here to stay. We must prepare for a future where they will comprise a large portion of business processes, the internet, and our daily lives. But to get there, we have to engineer them to greatness, and they have to earn our trust. It will require significant improvements and effort to reach its full potential. It will require high-quality context engineering and the full utilization of the MCP's potential. We’ve experienced the same in practice while building agents for enterprise customers. As Andrej Karpathy put it, “2025-2035 is the decade of agents”.
VP Engineering @ Mistral AI at the Parisian AI Engineer conference
There is immense potential, but you really have to find the right use case. We are especially impressed with Spotify’s talk about how they employ agents to upgrade dependencies across hundreds of repositories autonomously.
The process previously required extensive domain knowledge and involved numerous edge cases. Instead of writing complex logic to handle all of them, they wanted to create a simple prompt to instruct an agent. To verify the performance of this agent, they added robust verification steps, including compiling the code and running unit tests.
While they quickly saw the potential of this approach, there were also some drawbacks. For example, some of their tests were green because the agent had simply altered some unit tests to always succeed.
To solve this and many other minor problems along the way, the prompt grew and grew until it was the size of multiple pages. Ultimately, again, they were at a position where it takes a lot of domain knowledge (prompt engineering) to steer the migrations. However, the final agent solution can cover a significantly wider range of edge cases and, in turn, handle a larger portion of ‘the long tail’.
So agents are not a magical turnkey solution. It’s about building trust and reducing overhead systematically through multiple iterations, adding more checks and verifications throughout the agentic loop, and gradually granting more and more autonomy, and moving further down the long tail. As of now they are so confident in the solution that a lot of the PR’s that are opened by the agent, many of them are even merged automatically.
The event was hosted in a breathtaking venue. Take the Wintercircus in Ghent (which hosts about 35 tech startups) and do it times 100. Station F is the largest startup campus in the world. It hosts over 1000 startups annually. It is also a beautiful restoration of a historical building.
A term becoming more mainstream is context engineering. It’s the art of curating the input tokens to optimize the output tokens. Imagine a while ago, a user indicated they did not like to eat fish. If the current task or conversation involves preparing dishes, it's essential to bring this information to the model’s attention. The argument in favor of context engineering is that the current models are compelling, yet they mostly fail at tasks because they simply do not have the proper information to work with. Information that is often implicitly available in the minds of humans.
To unlock the full potential of these models, it is crucial to employ context engineering techniques. Some examples are dynamical tool discovery, context distillation, persistent memory creation and injection, or a specific one we learned: repeatedly prepending artificial safety instructions to enforce more adherence to company rules.
Daytona.io shared their vision of how the Agent Experience (AX) will enhance the user experience in a future where agents and assistants are the primary users of the internet (and thus your business applications). They envision that it is mandatory to start now with Agent-first app development: design applications that agents can easily and autonomously explore and use. To do so, design your application API first, provide an MCP server, and add llm.txt’s documentation. Also consider CLI-based solutions, as they align with how agents typically prefer to operate.
Agent-first development extends API-first approaches, and MCP will take a central role in this. Take the éas proof of this. It already has double the number of stars compared to the OpenAPI protocol repository in just a few months.
Kyutai is a non-profit organization to enables higher quality and personalized audio solutions with its breakthrough research. They talked about their low-latency audio solution Unmute.sh. They are also the makers of Moshi, the world's first full-duplex model, which can listen while talking to enable more realistic conversations.
Blackforestlabs has made significant progress on both the quality and speed of image generation by smartly utilizing the latent spaces. They not only allow you to generate images, but also smart edits such as reorienting the camera position, recoloring old photos, removing text or objects, and more.
Try to ignore the hype, but do keep trying out and building Agentic solutions yourself. It can drastically rethink business processes, but requires dedicated and long-running engineering effort to get it right. Prepare for a future with agents by starting to think about the Agent Experience and Agent-first development.
How to stay on top of what is happening in the Agentic world?