Introducing Agent Blueprint
Opening a conversation on agentic data infrastructure
Hello! Today, I’m launching this newsletter Agent Blueprint.
I’m Michel, co-founder & CEO of Airbyte. I’ve spent my career building infrastructure, first for data, now for AI. If there’s one lesson I’ve learned, it’s this: innovation stalls when there’s no standard. And in the past year, I’ve noticed something unsettling: AI feels magical on the surface, but messy underneath.
Every week I talk to founders, building agents and copilots. Their demos sparkle. But once you ask how it works? The story begins to fall apart. Agentic data pipelines are brittle. Data is stale. Permissions are bolted on at the last minute. Everyone hacks together the same pipelines without critically thinking about their data and AI strategy. Everyone struggles.
There’s no shared blueprint. No agreed-upon infrastructure. Just endless reinvention. Best practices vanish. And when every team solves the same problems in isolation, progress stalls.
Where we stand today with AI is exciting, but there is a lot of duct tape used. That realization is what pushed me to start Agent Blueprint.
Why I Started Agent Blueprint
We’re on the verge of a paradigm shift. Not just another wave of apps, but a transformation in how we build intelligence itself. And I believe the missing piece - what’s causing so many AI projects to underperform (only 5% of AI projects are successful today according to MIT) - is the definition of some standard in the AI infrastructure, and more specifically in the agentic data infrastructure.
Models impress us every day, they are not the bottleneck. They hallucinate because they lack the context that can properly meet our prompt’s needs and always try to satisfy us: cue the dreaded, “You are absolutely right.”
So what is the problem with the agentic data infrastructure? It’s the difference between agents that impress in demos (but fail in production) and those that scale, earn trust, and become foundational. And it is messy without standards. Even the most popular protocols, like MCP, are still far from being well-defined.
Context has many layers to it, from memory, to data acquisition from 1st-party tools (to potentially 3rd-party ones), to data transformation to make it actionable for the LLM or agent. One could say there are 8 layers to it, and there is no mature standard. That’s why I believe we’re on the cusp of a paradigm shift, and we now need a last mile of standardization to unlock the real potential of AI for our companies and society.
This newsletter, Agent Blueprint, is a project born of frustration, but above all from hope. The hope that by writing, thinking, and sharing openly, we can clarify what “context” really means in AI, and help accelerate the definition of those missing standards, and therefore the definition of the next generation of agents - agents that are reliable, grounded, permissioned, and built on infrastructures that don’t crumble when you scale.
Why Neutrality Matters
I want this to be a place where we make sense of the mess. Where we explore the paradigm shift AI is driving, but also acknowledge the gaps in infrastructure that hold it back.
I will be looking to approach it, not from the perspective of a vendor with something to sell, but as a neutral insider. I’ve had the privilege of working with thousands of teams through Airbyte, and I now sit at a vantage point that feels - to use a metaphor - a little like Switzerland. Close enough to see how the ecosystem is evolving, but neutral enough to speak honestly about what’s working, what’s not, and where we need to go.
My hope is this space becomes a shared foundation, a reference point when people ask: “What does context engineering even mean?”, “Which bits are solved, which bits are not?”, or “What is helpful vs what is hype?”
For instance, a few things I’ve seen:
Broken connectors everywhere: Notion, Slack, Drive, custom APIs. Every team building agents spends at least weeks (often longer) just getting data in, dealing with format mismatch, permissions, stability.
Permissioning as afterthoughts: either people skip them, or build naïve versions that fail.
Prompt engineering workarounds, not solutions: hacks, temperature tricks, preprocessing. These are necessary, but brittle and inconsistent.
Control & sovereignty overlooked: too many teams rely on closed platforms without thinking about data ownership, portability, or the ability to switch providers. This creates hidden lock-in and limits the freedom to innovate.
Enterprise trust is fragile: security, compliance are often retrofitted or ignored until disaster. And often, by then, the ship has sailed.
So I asked myself: what would the ecosystem look like if we had standards, a shared understanding, for context engineering? If someone writing agents didn’t need to think from scratch about connectors, chunking, metadata, permissions, they could move faster, invent more, experiment more safely. Everybody wins. That belief is what fuels Agent Blueprint.
What Agent Blueprint Will Be About
Here’s what you’ll get, and what I plan to deliver, week after week:
Framework Essays: deep dives into what the stack looks like - e.g. “The 8 Layers of Context Engineering” - to help you map your architecture and spot weaknesses.
Field Reports & Interviews: lessons from AI founders building in the mess - what works, what doesn’t, what they wished they’d known.
Technical Explainers: breakdowns of data pipelines, ACL systems, cost/latency optimization, and how to build systems that scale.
Category Thinking: thoughts on new emerging categories, exploring neutrality, debating what standards look like, how protocols evolve.
One promise: this will always be free. Knowledge-sharing should accelerate the ecosystem, not gate it.
Always free, but also always honest. Always trying to pull back the curtain on what works, what’s aspirational, and what is dangerously misleading if you accept it uncritically.
What You Can Expect From Me, & Me From You
If you subscribe:
You’ll get essays like this - personal, technical and strategic.
You’ll see frameworks, predictions, and also disagreements: sometimes I’ll be wrong; I want your feedback.
I’ll share lessons and mistakes publicly - not polished case studies only.
If you’re building agents, copilots, or context-heavy AI systems, I hope you’ll see Agent Blueprint as useful. If something here resonates, share it. If something doesn’t, tell me.
If you believe that what agents need most today is not just better models, but better infrastructure, better standards, better definitions of context, better discipline, then let’s walk this path together.
Subscribe, join the conversation, and let’s build the missing blueprint for the next wave of AI.
Michel