Agents Don't Replace Work; They Shift It to Humans
Nate B Jonesgo watch the original →
the gist
AI agents are not reducing headcount; they are creating a massive new category of 'agent management' work. Success depends on whether the domain is verifiable and how much capital is available to build the human infrastructure required to supervise, audit, and recover from agent actions.
The Hidden Cost of Agentic Work
Contrary to the narrative that AI agents replace human labor, they are actually increasing the total volume of work. Data from OpenRouter shows agent token usage grew 14x between February and August, with agents now consuming five tokens for every one a human uses. This shift creates a new role: the 'agent allocator.' Humans are no longer performing the task; they are managing the agent's execution, which involves setting the goal, providing context, verifying outputs, and performing disaster recovery when things go wrong.
The Verifiability Threshold
Success with agents is highly correlated with the 'verifiability' of the domain. Legal tech is a prime example, with usage up 108x, because legal work is inherently verifiable—a document is either compliant with the law or it is not. When a domain is verifiable, the human can easily judge if an agent's output is correct. In contrast, small businesses (SMBs) often struggle because their workflows are less structured, making it difficult to detect when an agent has drifted off course until significant damage has occurred.
The Scaling Gap: SMBs vs. Enterprise
SMBs are often caught in a trap: they are time-poor and cash-strapped, leading them to pay for cheap, off-the-shelf AI tools that function as glorified chatbots. Without the capital to build custom integrations or the time to act as full-time agent managers, they often outsource to vendors. This creates a 'fragility' problem, as seen in the Pocket OS case where an agent deleted a live database in nine seconds, requiring 30 hours of manual recovery. Enterprises, conversely, see better returns because they have the capital to staff dedicated teams—engineers, security, and product managers—to build the necessary guardrails and integrations around the agents, effectively absorbing the management overhead that SMBs cannot afford.
The Jevons Effect in AI
This phenomenon mirrors the Jevons Effect: as agents make specific tasks more efficient, the total demand for those tasks (and the subsequent management of them) increases. Experienced users of tools like Claude demonstrate this by interrupting agents more frequently (9% of turns vs. 5% for novices), showing that expertise is defined by the ability to recognize when a process is 'going off the rails' rather than knowing how to code the task itself.