AI is not going to replace humans. It is going to replace human-judgment-based workflows. Someone still has to supervise. Someone has to notice when an agent is drifting. Someone has to approve changes. The shape of work changes; the need for judgment does not.
The difference is scale. A manager handles about seven direct reports — the limit because humans are emotionally complex and require constant context-switching. Agents are different. Most of them, most of the time, are just cruising along. A human should be able to supervise dozens or hundreds, if the observability and review layers are good enough. The atomic mesh is the only architecture that scales human oversight, because it is the only architecture where you can localize the thing that needs attention.
The missing piece is outside stimulus. Humans self-correct through conversations, feedback, social cues. Agents have none of this by default. A supervisor layer — human, automated, or hybrid — has to provide that stimulus deliberately. A higher-level agent reviewing subordinate outputs on a cycle. A human reviewing a dashboard weekly. Automated tests flagging drift. Some feedback loop must exist, because agents without outside stimulus do not learn. This is the operational realization of non-determinism demanding human correction loops — the loop has to be explicit because it does not happen for free.
The teams that get this right will look weirder than today's org charts. One supervisor accountable for fifty atomic agents that ship work into different parts of the business. A small bench of reviewers who specialize in artifact types, not departments. The supervisor's job is less management and more dashboard reading, drift detection, and periodic correction. It is a new discipline.
If you are designing an agent rollout, plan for the supervisor role explicitly. Who watches the agents? What surface do they watch them through? How fast can they intervene when something is off? Without those answers, the fleet does not scale — no matter how good the individual agents are.
Sources
Related Essays
The Primitives Are the Same Across Roles
Sandbox, tools, prompts, governance — the primitives are universal. Engineering agents and GTM agents share the same engine. Only the body changes.
Agents Are About to Break Out of Engineering
Non-engineering agents are easier to build but harder to contextualize. The organizational knowledge layer — not the execution layer — is the real unsolved problem.
The Grayscale Between Engineering and Everywhere Else
There is no clean split between coding tools and business tools. The reality is a grayscale, and the products that win serve the whole spectrum.
Key takeaways
- Human management caps at about seven direct reports because humans are emotionally complex.
- Agents are different. Most of them, most of the time, are cruising. A human can supervise hundreds.
- Without outside stimulus, agents do not learn. A supervisor layer must provide that stimulus deliberately.
FAQ
Why is the human cap so different for agents?
Because agents do not require constant context-switching the way people do. Most agents do not need attention most of the time. The bottleneck moves to observability and review surface quality.
What is the supervisor layer?
Whatever provides outside stimulus to agents. Could be a higher-level agent reviewing subordinate outputs on a cycle. Could be a human reviewing a dashboard weekly. Could be automated tests flagging drift. Some feedback loop must exist.