You don’t have a model problem. You have a harness problem.
The model is about a tenth of a working agent. The other nine tenths is the thing your team has to build, and in data, almost nobody is building it.
Every time an AI tool disappoints a data team, the conversation goes the same way. Wrong model. Wait for the next one. It’s almost never the model.
Google’s number for this is blunt: in a working coding agent, the model is roughly a tenth of the system. The other nine tenths: the tools, the context, the sandboxes, the orchestration, the guardrails, the observability. Google calls that the harness. The equation they use is Agent = Model + Harness, and the evidence behind it is hard to argue with. One team moved a coding agent from outside the top 30 on a public benchmark into the top 5 by changing only the harness. No model change at all.
Their conclusion is the part worth pinning above your desk. Most agent failures, examined honestly, are configuration failures. And the harness is the team’s surface area, not the model provider’s.
What the harness is, in data
Take their component list and ask what each one means when the thing being produced is a number rather than a function. It maps almost too neatly.
Instructions and rules become the semantic model: concepts, grain, exclusions, the join decisions a person ruled on. Tools become your estate: the catalogue, the warehouse, the workbooks and reports where the real rules have been living. The sandbox becomes a design plane that runs with nothing connected. Orchestration becomes the lifecycle, each stage refusing entry until the previous one is honest. Guardrails become controls bound at design time and a release token the orchestrator has to read. Observability becomes evidence: the control version, the parameters, the query, the observed value.
Nobody in data is short of models. Everybody in data is short of harness.
Why this is the whole strategic point
Because it tells you where to spend. If the model is a tenth of the outcome, then picking a vendor by benchmark score is optimising the smallest term in the equation. The nine tenths that decide whether this works on your estate is something you either build, buy, or quietly do without, and most organisations are doing without while blaming the tenth.
It also explains the pilot graveyard. The text-to-SQL pilot that impressed everyone in the demo and died in month three didn’t die because the model was weak. It died with no stated grain to constrain it, no controls to catch it, no evidence to defend it, and no sandbox to design in before the credentials arrived. A raw model with no harness isn’t an agent. It’s a very fast guess.
How to spend like you believe it
We didn’t set out to build a harness for data agents. We set out to make meaning survive a hand-off between people. They turn out to be the same object, which is either a coincidence or a sign that the problem was never really about AI.