Agent harnesses: model makers' own stacks lead; no shared standard within a year — resolves 2027-07-14

The debate: should the harness, the tooling wrapped around a model that lets AI actually carry out tasks, be built vertically by each model maker, or unified by a vendor-neutral shared layer that can swap models, of the type Databricks' Omnigent represents?

Why it matters: whoever owns this layer decides where a company's AI workflows are tied down and how easy it is to change the underlying model.

Who is on each side: model makers' own vertical stacks lead by a wide margin in usage today, while backers of a shared layer point out that users have already started swapping the underlying model on price and performance.

SecondSource’s reading: in the harness layer, the tooling wrapped around a model so AI can actually carry out tasks, vertical stacks built by the model makers themselves lead today, by a wide margin in usage.

But the habit of swapping the underlying model on price and performance has already appeared, an early sign for the meta-harness route, a vendor-neutral shared layer that can change models.

We expect that within 12 months such a shared layer will not yet become the de facto standard.

Ethan Mollick (Wharton professor, One Useful Thing)
“Because of this change, you have to consider three…”

Settled over 12 months by whether a meta-harness becomes the de facto standard, evidenced by at least two large enterprises saying so publicly. If it does not, the call stands.

What would overturn it

A common layer of the Omnigent type becoming the de facto standard.