“Good tools are cached intelligence for agents.”
While the AI scene tends to orbit around the same few names, Hugging Face quietly builds, in the background, a large part of the infrastructure the whole ecosystem runs on.
What the team published last week is—on the surface—a redesigned CLI, made agent-aware. It detects whether it's being called by a human or an agent, adjusts its output accordingly, embeds workflow hints to reduce back-and-forth. Coding agents using it consumed 1.3× to 6× fewer tokens than with curl or SDK alternatives, at higher success rates.
But Clem's sentence points at something deeper than benchmarks, and it sent me back to a much older story—one Arthur Mensch outlined from a different angle in his much-noticed hearing before the Assemblée Nationale a few weeks ago.
We are living through the power plant phase. The benchmark races, the frontier lab announcements, the parameter counts—all of it keeps attention on the generation layer, on who produces the most intelligence, on which model reasons best. The same debate played out a century ago—about coal output, hydroelectric capacity, and who was building the biggest plant.
Essential—but not the whole answer.
As electricity became abundant—cheaper, more reliable, more local—part of the economic value migrated: first into motors and the machines around them, then into industrial processes, and eventually into the software and platforms on top. The kWh stayed essential throughout. It captured a fraction of what it helped create.
The token may be heading somewhere similar. When intelligence becomes genuinely cheap—locally available, continuously flowing—the question that matters shifts. Not "how expensive is reasoning?" but "how efficiently do we convert it into useful work?"
Which is what makes Clem's framing quietly radical. A good tool doesn't just wrap a function: it compresses thousands of hours of domain reasoning into a single call. The workflow is already designed, the abstractions already there, the failure modes already handled. When an agent invokes that tool, it inherits all of it—without spending a token on rediscovery. That's not a convenience. It's a gearbox—and gearboxes don't generate energy, they convert it.
This is what harness engineering is becoming: not smarter models, but better systems around them—skills, MCPs, memory, orchestration, and the right model for the task. Specialized models like Mistral's start looking less like a "smaller frontier" and more like mini mobile power plants: local, ultra-efficient, deployable wherever the work actually is.
The teams that internalize this early—that ask not "which model is most powerful?" but "how do we convert intelligence most efficiently?"—may end up in a strong position. Not because they processed the most. Because they learned to waste the least.
