Summary
Despite widespread expectations that AI agents would reduce headcount and free up human time, early usage data tells a more complicated story: agents are generating significantly more supervisory work for humans, not less. Nate B. Jones draws on Open Router data showing agent token consumption grew 14-fold between February and August 2026, with agents now burning more than five tokens for every one produced by a human. OpenAI reports that its heaviest Codex users generate over 60 hours of agent activity per day — far beyond what any individual can meaningfully monitor.
The video examines how this dynamic diverges sharply across company sizes. Small and medium businesses tend to struggle: most operate in domains where it’s hard to verify whether the agent did its job correctly, they lack the capital for deep integration, and vendor-managed agents often create value for the vendor without transferring lasting capability to the business. Legal is a notable exception — Codex usage in legal contexts has grown roughly 108x since January 2026, driven by law’s inherently verifiable structure. Enterprise deployments, by contrast, tend to report better returns because larger companies can afford dedicated engineering teams to build agent management into their workflows.
Jones argues that someone still has to choose which jobs to run, validate outputs, and intervene when agents fail — and that “agent manager” is itself a new and growing category of work. The central question is whether management overhead will grow faster or slower than the productivity gains agents actually deliver as the technology matures.
📺 Source: AI News & Strategy Daily | Nate B Jones · Published August 26, 2026
🏷️ Format: Opinion Editorial
