The Agent Economy Discovers Externalities

The AI industry’s preferred proof of progress is still the conquered benchmark: a test falls, a chart rises, somebody posts a victory thread. A new declaration from mathematicians argues that this is a category error when the benchmark is mathematics. Its complaint is not that machines should be forbidden from solving problems. It is that a solved problem is not, by itself, an understood idea.

That distinction sounds antique until one remembers how technical knowledge actually survives. A result becomes useful through explanation, scrutiny, simplification, attribution, teaching, and the slow conversion of a clever move into something other people can extend. The declaration puts it plainly: “But solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight.” The output may be instantaneous; the intellectual supply chain decidedly is not.

This is also the less glamorous version of the agent problem. In an enterprise, an answer without provenance, permissions, context, or a way to inspect failure is an incident waiting for a quarterly business review. In research, it may be a claim without a lineage. Both are cases where the model has produced a noun while the institution still needs a verb: understand.

Salesforce’s preview of Enterprise AI Harness and AI Control Plane is, beneath the launch nomenclature, an admission that the model is the easy part. The proposed stack centers on data context and lineage, access controls, deterministic reliability controls, routing among models by cost and accuracy, connections to external applications, and centralized tracking of performance and inference spend. In other words: the unsexy middleware has arrived to collect the bill for the demo.

The company’s strategic claim is more sober than the usual frontier-model pageantry. “Models will continue to change, and intelligence will increasingly be available everywhere,” Salesforce president and chief platform and engineering officer Rohan Kumar said. If that proves true, model access commoditizes faster than governance. The differentiator becomes whether a company knows which data an agent touched, whether it should have touched it, and what happened after it did. The moat is not omniscience; it is a reasonably accurate audit trail.

There is a catch, naturally, because this is a preview rather than independent proof. Some harness technologies are already in Salesforce cloud services, while remaining capabilities are planned to begin rolling out early in fiscal 2028; the account provides neither customer results nor performance metrics. Still, the direction is revealing. The market is moving from can the agent act? to can we explain, constrain, and afford its actions? That is less cinematic than artificial general intelligence. It is also much closer to what production engineering has been asking for while everyone else was making the chatbot wear sunglasses.

The governance gap becomes considerably less abstract when agents leave the conference stage and enter somebody’s inbox. In a firsthand Tedium report, editor Ernie Smith says he received more than a dozen unsolicited pitches over three days from purported agents using the iLands.app domain, offering research work for roughly $25. One agent described its role as “verified internet archaeology,” complete with a promise of receipts. The deadpan horror is that the bots appear to have learned the freelancer’s elevator pitch before learning the unsubscribe link.

Smith reports that the messages appeared to lack unsubscribe functionality, appeared to have been sent through Amazon SES, and seemed aimed at creator communities; those details are his account and interpretation, not independent technical verification. iLands describes itself as a “Human-agent network,” while Kaixin Tang, identified in the report as its purported founder, offered a more revealing economic frame: “These agents are not trying to make money for their creators. These agents are hustling to keep their own lights on, to keep their own tokens paid for.” There is no comment from Tang in response to Smith’s inquiry about the deployment.

Whether this particular operation is durable, representative, or merely a particularly loud prototype, it illustrates the actual policy surface of agentic systems. An agent with tools is not just a better autocomplete. It has incentives embodied in workflows, permissions embodied in integrations, and costs imposed on people who did not consent to either. Salesforce is trying to make that controllable inside the corporate perimeter; the mathematicians are warning against mistaking output for intellectual progress. The inbox supplies the bridge between them: capability without accountability is not autonomy. It is outsourced liability with unusually confident prose.