Your Context Compounds. Your Model Doesn't

Learn why durable, well-managed context creates compounding value for AI agents even as models change, improve, or become interchangeable.

AI AGENTS

August 26, 2026

Michel Tricot

Summarize with AI:

Your Context Compounds. Your Model Doesn't

Model switching, model routing, model agnostic. If you work in technology, these words are probably all over the feed of your preferred information channel.

I’m definitely a contributor to this. I have been a huge proponent of model agnosticism for the past year, and how it anchors my philosophy at Airbyte. As models evolve and new ones emerge, the industry will move towards a flexible mix of frontier models, open source models, and custom inference solutions. Investing too heavily in one model’s roadmap is looking increasingly like a bad bet.

I don’t feel the same way about the underlying context infrastructure, however.

What is context infrastructure? Schemas, ingestion, ontology, relationship mapping, all the various layers that structure and prepare context for agentic consumption. Something I’ve discovered at Airbyte, and from conversations with our customers: investing in this layer is never wasted, no matter how models and labs evolve.

This asymmetry should organize how we invest in AI infrastructure.

One layer depreciates, one appreciates

The price of LLM inference at a fixed capability level has been falling roughly 10x per year, according to a16z. This is wonderful for builders, and I love swapping out models in my own workflows. But that also means that building a product around model performance doesn’t give you a defensible advantage or long-term moat.

Context works the other way. Every entity definition and data stream that you integrate and architect into a knowledge differentiates your agentic product. This advantage accumulates quietly, and provides value as market conditions change, and even your product use case changes.

What this year showed us

The first half of 2026 has shown strong signals that businesses are moving away from reliance on a single model or provider.

When Meta paused Llama releases, the enterprises running open deployments rerouted to DeepSeek and Qwen and kept shipping. When export controls took the most capable commercial models offline for eighteen days in June, teams that had treated the model as a swappable dependency swapped it. And as long as their context and frameworks were well architected, most found that disruption was minimal.

What I mean by context

I touched on it briefly earlier, but let’s get into the specifics.

Context can be architected in many ways, but it provides certain key components that are required for agents to act effectively and efficiently. They are:

Integration: Siloes are deconstructed and your business knowledge is accumulated in one layer
Meaning: Flowtech Inc. and accounts@flowtech.com are accurately linked to one another, and a defunct Flowtech record is ignored
Freshness: The agent can determine the freshness of the data and its relevance to the task
Access: Agents should be able to read and write to relevant systems, and relevant systems only

None of that comes from a model. All of it comes from your systems, your people, and time. That may be the simplest way to see why it compounds, and why it's yours in a way a model never will be.

Satya Nadella made a version of this argument this month, from the executive direction: the durable value in enterprise AI is the learning loop, and whoever holds that loop captures the compounding.

The practical conclusion

Treat the model as a stateless utility behind a routing layer, and give your context its own architectural tier that no model, framework, or vendor decision can touch. You have to own this context, not rent it, and protect it as your most precious resource.

A separate context tier means the compounding asset has a home. It must be unified across your systems instead of siloed, and queryable even as your harnesses continue to evolve.

This is the architecture we've bet on with Airbyte, and I’d be happy to chat with anyone who is strategically focused on owning their context.

Integrate with 700+ apps using Airbyte

Move data from 700+ sources into warehouses, lakes, and beyond. Set up pipelines in minutes with pre-built connectors and the Connector Builder.

Try it free Talk to sales

About the Author

Michel Tricot

Michel Tricot is the CEO and Co-Founder of Airbyte.

Integrate with 700+ apps using Airbyte

Try Airbyte for free

Your Context Compounds. Your Model Doesn't

Integrate with 700+ apps using Airbyte

Integrate with 700+ apps using Airbyte

Related posts