AI has dramatically increased the speed at which software can be built. A developer can describe a feature, have an AI generate hundreds of lines of code, test an approach, correct an error, and move to the next task in a fraction of the time the same work once required.
But something interesting happens as that capability increases.
The AI gets faster. The human does not.
That may become one of the most important constraints in AI-assisted development.
The Human-Attention Bottleneck
Most AI development tools still use a workflow inherited from traditional software development: the AI performs some work and explains what it did, the human reviews the explanation and the changes, makes another decision, and the process repeats.
That makes sense when work is relatively slow. It becomes inefficient when an AI can perform dozens of technical operations in the time it takes a person to properly understand one of them.
The result is a strange inversion. The machine may be capable of doing more work, but the operator increasingly spends time watching the machine work.
More Output Does Not Necessarily Mean More Progress
I noticed this while building increasingly complex systems with AI. The amount of work being produced went up substantially, but my ability to maintain a coherent understanding of everything happening did not keep pace.
There was more technical narration, more changes to inspect, more branches of work, and more moments where I had to decide whether something mattered enough to interrupt what I was already thinking about.
The AI was increasing activity throughput, but that did not always translate into completion throughput. A system can generate ten times more work, but if that creates ten times more things for the operator to track, review, remember, and reconnect later, much of the leverage disappears.
We may be measuring the wrong thing. The important metric may not be how much code the AI can generate, but how many completed objectives we can achieve per unit of human attention.
Attention Does Not Scale Like Compute
Compute can be added. Agents can be added. Human attention cannot scale the same way. A person still has to understand consequential decisions, maintain strategic direction, resolve ambiguity, and recognize when something is moving in the wrong direction.
AI can increase production faster than a human can increase attention.
At first, the limitation was how quickly we could write code. AI largely removes that constraint. The next limitation is how much generated work a human can meaningfully supervise.
We Probably Shouldn't Review Everything
Better summaries help, but I don't think summarization alone solves the problem. The deeper question is why the human needs to see routine execution at all.
A CEO does not sit beside an accountant and watch every transaction get entered. A project manager does not observe every keystroke made by an engineer. We establish objectives, procedures, constraints, verification, and escalation rules. Competent work happens inside those boundaries, and attention is requested when something falls outside them.
AI systems will likely need to evolve the same way: supervision by exception. Routine work executes inside defined constraints, the system maintains the complete technical record, verification happens automatically wherever possible, and the operator primarily sees states such as Completed, Verified, Blocked, Conflict detected, and Decision required.
The technical detail should still exist and remain inspectable. It simply should not automatically consume the operator's attention.
Work Needs Containment
Today, much of an AI development session happens in one continuous conversational stream, where requirements, implementation details, debugging, architecture, status updates, questions, and decisions all compete for space. That works for small tasks. It breaks down as the scale of AI-directed work grows.
A better model is containment:
- Workers contain their work.
- Tasks contain their records.
- Projects contain their state.
- The operator's conversation contains only what deserves the operator's attention.
The detailed work does not disappear. It moves to the level where it belongs, and the human interface stays focused on decisions, objectives, exceptions, and consequences.
AI Changes the Meaning of Management
This is not only a software-development problem. If every AI agent continuously reports everything it does to a person, adding more agents eventually makes the human less productive rather than more productive.
A useful AI system should not require the human to become the communication bus connecting all of its parts. The system itself needs to preserve state, coordinate work, enforce procedures, record evidence, and determine when human judgment is actually required.
The role of the human moves upward: less attention spent watching execution, more on architecture, intent, principles, priorities, and consequential judgment.
The Next Productivity Gain
The first major productivity gain from generative AI has been obvious: machines can produce intellectual work much faster.
The next gain may be less obvious: systems that allow humans to consume less attention while directing more capability. That means better containment, verification, state management, and escalation, and better protection of human cognitive bandwidth.
The most capable AI system may not be the one that tells us the most about what it is doing. It may be the one that can reliably determine what we actually need to know.
The limiting resource may no longer be code. It may be attention.