Main idea

There is a precedent: as long as SaaS products lived in isolation, they were compared on the richness of their features. The day they had to coexist inside the same organization, the API became the criterion that eliminates — a tool without an API, however good, costs more in re-entry than it gains in functionality.

The same shift is replaying with agents. A tool whose agent sees only its own data produces output that is consistent with one face of reality and wrong about the rest. A context exchange protocol — MCP being the form that is emerging — lets Figma go and read the recent changes in a GitLab repository, understand the relevant merge requests, and align the design with what the product has become.

Each tool's specialty doesn't disappear; it stops being what decides. What decides becomes the ability to retrieve context elsewhere and act across a perimeter wider than its own.

Layer added by "Running my Contexts directly inside ChatGPT" (2026-09-20). The shift can already be observed on consumer tools, and it is measured by the number of gestures it removes. A conversation recorder reachable through a connector lets you ask, from ChatGPT, for the list of the latest recordings and the direct retrieval of their transcripts. Without a connector, the same material costs five manual operations — download, open, copy, go back, paste — each of which rules out use on the move. So it is not the transcription feature that has value, it is that it is reachable from the tool where the work happens.

Why it matters

This gives a tool-selection criterion that comes before use: assess what a tool can read outside itself, before assessing what it can do. The criterion has already proved its predictive value once, with APIs.

It also says what happens to established positions. A tool that dominates through its functional perimeter becomes vulnerable as soon as the comparison bears on openness.

Nuances and limits

The analogy with APIs has a weakness: an API exposes data whose format is known and stable, a context exchange protocol exposes material that a model interprets. Openness therefore does not guarantee that what is retrieved is understood.

And the movement is not complete. The mechanism is visible in its direction, not in its final equilibrium, and nothing guarantees that the protocol which prevails will be this one.

Open questions

• What decides, within an organization, which tool becomes the one that retrieves and which becomes the one that supplies?