Reading Claude as a Full Time Job

Inside one of the world's most AI-intensive companies, employees are becoming managers of AI agents. Its latest research offers a fascinating glimpse into what the future of management might look like.

By Sharon Gai

We have spent more than a century developing sophisticated theories about how to organize human labor. Frederick Taylor studied how to make industrial workers more efficient. Henri Fayol developed principles of administration and organizational control. Peter Drucker helped us understand how to manage knowledge workers, whose contributions could not be measured simply by counting the number of things they produced.

But almost all these theories share one assumption: the workers being managed are human.

That assumption is beginning to change.

I have been spending a lot of time thinking about what it means to manage AI agents. We have spent more than a century developing theories about how to manage people, but what happens when the work is increasingly performed by software? How many agents can one person realistically supervise? What should we delegate to them? And how do we know whether the agents are actually making the organization more productive?

A new study from Anthropic gives us one of the clearest pictures yet of what happens when a company starts reorganizing itself around AI.

On October 16, Anthropic published a report called How Employment Is Changing at Anthropic. Since Anthropic is both developing some of the world's most capable AI models and using them extensively within its own operations, it offers an interesting preview of what other organizations might eventually experience.

The headline finding is remarkable. By August 2026, AI agents directed by Anthropic employees were collectively performing six hours of active work for every estimated hour worked by a human employee.

Now, before we get too excited, this doesn't mean Anthropic has become six times more productive. The researchers measured agent runtime, including inference and active tool use, and compared it with human working hours estimated using a forty-hour workweek. These are not economically equivalent units. An agent could spend an hour attempting something that a human could complete in ten minutes, or produce an output that requires substantial correction.

But even with that qualification, the finding is extraordinary. We are looking at a company where the amount of machine execution has already surpassed the amount of human working time by a considerable margin.

What interests me even more is what the humans are doing.

Anthropic Figure 2. Agent hours overtake human hours. Employee-directed interactive sessions, including pooled subagents. Runtime counts inference and active tool use, excluding waits longer than 30 minutes. Human hours assume a 40-hour workweek.

Source: Anthropic, October 16, 2026. Research report · Original figure

The use of AI curiously increased meetings  

You would think that by using AI in general would increase efficiency and thus reduce meetings right? The median employee was spending approximately 20% more time in meetings than a year earlier.

As agents take over routine execution, humans may spend a greater proportion of their time discussing strategy, resolving ambiguity, and making decisions that cannot easily be delegated. Employees may write less code or produce fewer documents directly, but still need to coordinate the consequential decisions that guide all that machine-generated work.

Anthropic also found that employees were under pressure to make their work more visible to agents. A decision made in a private meeting, for example, might never reach the systems that need that information. An agent could continue operating on outdated assumptions simply because nobody documented what had changed.

This creates a new organizational challenge. Companies will need to think carefully about how information flows between people and agents, which decisions are recorded, and how agents obtain reliable context.

The traditional organizational chart tells us who reports to whom. But it doesn't explain which agents perform which tasks, who supervises their decisions, or how information moves between humans and machines.

We may eventually need a second map of the organization, one that shows how work is orchestrated across people, agents, and enterprise systems.

This brings me back to an idea I often share in my book: stop being the busy bee, become the beekeeper. In a traditional pyramid, layers of people relay information upward and instructions downward. Authority is tied partly to what a manager can see and pass along. In the hive I have in mind, agents help route information, report on progress, and draft the work. Humans set direction, decide what matters, and make the consequential calls. The work moves through a network of people and agents, with shared context helping them coordinate without every update climbing a reporting ladder.

From pyramid to hive. My beekeeper metaphor for organizational design: AI agents help route, report, and draft, while humans set direction and make the call.

Source: Sharon Gai, supplied illustration.

Becoming the beekeeper means designing the conditions in which that network can work responsibly. Someone still has to define the goal, decide what agents may do independently, and establish which decisions require human review. The beekeeper watches the health of the whole hive, intervenes when something goes wrong, and remains accountable for the outcome. My pyramid-to-hive illustration is a way of imagining that organizational shift, rather than a claim that Anthropic has already replaced its hierarchy. It connects directly to attention-based span of control: a hive can grow only if its demands for human judgment remain manageable. Funnily enough, in an essay by Steve Yegge, Anthropic was actually described as a hive.

I think this could become one of the defining challenges of organizational design in the AI era.

Management is becoming less about supervising people and more about coordinating the execution of work

At Anthropic, employees are using agents to write code, investigate bugs, conduct research, manage customer information, review legal documents, and perform administrative tasks. Some of these activities previously required employees to complete the work themselves. Now, employees increasingly spend their time defining what needs to be done, instructing agents, evaluating their results, and deciding what happens next.

One of my favorite examples from the report involves a team responsible for writing content that appears in Claude's interface. Think of the little messages and greetings users see when interacting with the product. Traditionally, this was the responsibility of human writers. But in spring 2026, the team developed a system that allowed agents to generate much of that content.

The writers then moved into a different kind of role. Instead of writing individual messages, they began designing prompts, building evaluations, and improving the instructions that guided the agents.

I find this transition fascinating because the employees were no longer simply writers. They had become managers of a writing system.

This is an important distinction. In most organizations, we associate management with seniority. Someone begins their career by doing the work, develops expertise, and eventually gets promoted into a position where they supervise others. But AI agents could change that progression. An individual contributor might now direct ten or twenty agents without having a single human direct report.

And that raises a question I've been exploring in my work on Agent Management Theory: If an employee spends most of the day delegating tasks to agents, reviewing their performance, and improving their instructions, aren't they already performing management work?

Anthropic's research suggests that this is happening at scale. In an employee survey, one in five respondents described their role shifting from directly performing work toward managing Claude. The researchers also found that the proportion of sampled interactive Claude sessions devoted to what they call meta-work, such as building agent instructions, monitoring systems, and coordinating agent activities, increased from 10% in fall 2025 to 22% in spring 2026.

In other words, employees are increasingly using AI to manage AI.

One engineer even created an agent chief of staff to help triage requests from other agents. Another wrote something resembling a performance review for their Claude team, identifying what was and wasn't working in their communication.

We have apparently reached the point where people are conducting performance reviews with software.

It sounds amusing, but I think there is something important happening here. Management is becoming less about supervising people and more about coordinating the execution of work, regardless of whether that work is performed by humans or machines.

We are becoming babysitters, babysitting agents

Here is where the story gets particularly interesting.

Despite the enormous amount of work Anthropic's agents are performing, the researchers discovered that interactive agents spent approximately 80% of their session time waiting for human input.

Think about that for a moment. We have built machines capable of working continuously, and yet much of their elapsed session time is spent waiting for us.

One engineer compared the experience to babysitting, where every child demands attention with the same urgency, regardless of whether the problem is trivial or serious.

I can imagine how this happens. Suppose you have ten coding agents working on different projects. One needs permission to access a file. Another encounters an unexpected error. A third wants clarification about a design decision. A fourth has completed its assignment and needs you to review the result.

Individually, these interruptions may be manageable. Collectively, they can consume an enormous amount of attention.

I see this with my own work. Sometimes an entire workflow is blocked because I accidentally forgot to click "grant permission" to this particular Claude thread who needed to use some sort of tool to accomplish a task I gave it. In many steps of the way, the agent needed me to grant access, provide context, and I became the bottleneck.

But this is the autonomy struggle. When we let our agents loose, and we forgo supervision, it can also go rogue. We've seen so many cases this summer of what happened when we let our agents run free. So it's a balancing act. And one that experienced managers will know better than junior employees - deciding how much "freedom" or micromanagement you want to give to your agents. Give too much, and they might run wild. Give too little, and you risk slowing down.

This reminds me of a familiar problem in growing companies. A founder hires more employees because there is too much work to do, but eventually discovers that everyone is waiting for the founder to approve decisions. The company has increased its execution capacity without increasing its decision-making capacity.

AI agents could create the same bottleneck, except at a much greater scale.

The Anthropic researchers found that even nontechnical employees were generating enough interactive Claude output that reading all of it would consume more than half their working week. That is an estimate based on output volume and reading speed, not evidence that employees actually read everything. Still, it illustrates the scale of the information problem.

Anthropic Figure 6  Reading Claude as a full-time job. Estimated time to read interactive agent output, using 0.75 words per token and 250 words per minute. This measures output volume, not time employees actually spent reading.

Source: Anthropic, October 16, 2026. Research report · Original figure

If your agents can produce more work than you can possibly review, the limiting factor is no longer how quickly the work can be generated. It is how quickly a human can determine whether the work is useful, correct, and worth acting upon.

This makes me wonder whether we need to rethink the traditional management concept of span of control.

For decades, management theorists have debated the ideal span of control, or how many employees should report to one manager. Traditional management structures often use around five to ten direct reports as a reference point, although the appropriate number varies considerably depending on the complexity of the work. A manager overseeing highly specialized professionals may have only a handful of direct reports, while someone managing a standardized operational process could supervise twenty or thirty people.

But what happens when those direct reports are AI agents? Is the right number ten, fifty, or five hundred?

I suspect the answer has less to do with the number of agents than with the amount of attention they require. A manager overseeing fifty reliable agents that perform routine tasks might have a relatively easy job, while another manager overseeing three agents making complicated financial decisions could spend the entire day resolving exceptions and reviewing their outputs. Perhaps the more useful measure is not how many agents a person manages, but how much productive work those agents can complete per hour of human supervision.

This is one of the first concepts I would want to study in developing Agent Management Theory. We may need to move from a headcount-based span of control to an attention-based span of control, where managerial capacity is determined by the frequency, complexity, and consequences of the decisions that require human intervention.

Output per Employee Rose Across Technical and Business Roles

Anthropic also found evidence of substantial productivity improvements. Code commits per engineer increased, sales representatives handled more customer activity, and employees reported automating work that previously consumed significant amounts of time.

One account executive described automations that eliminated approximately 20% of their previous tasks. Another said automation allowed them to spend more time with customers and engage with approximately 30% more of them.

These are meaningful examples because they show how agents can change the allocation of human effort. The sales representative isn't simply sending more emails. They are using automation to spend more time on customer relationships.

Anthropic Figure 3  Productivity boom across functions. Smoothed output per person, indexed to December 2025. These descriptive trends do not by themselves establish how much of the increase was caused by AI.

Source: Anthropic, October 16, 2026. Research report · Original figure

But I think we need to be careful about how we interpret productivity claims.

An increase in code commits does not automatically mean an equivalent increase in valuable software. A greater number of customer interactions does not necessarily produce higher revenue. And more agent activity certainly does not guarantee greater profitability.

What matters is whether the combination of humans and agents produces better business outcomes after accounting for the costs of operating and supervising the agents.

Imagine an engineer who previously wrote five hundred lines of code a day. With coding agents, that engineer can now generate five thousand. On paper, productivity has increased tenfold. But what if the engineer spends the rest of the week reviewing the code, fixing errors, resolving integration problems, and investigating unexpected behavior?

The organization has produced more code, but has it produced more value?

The same problem appears in marketing. An agent can generate fifty campaign concepts in the time it once took an employee to create five. But someone still needs to determine which ideas are good, whether they align with the brand, and whether they will actually influence customer behavior. More output can simply create more decisions.

This reminds me of Drucker's distinction between efficiency and effectiveness. Efficiency is about doing things right. Effectiveness is about doing the right things. AI agents may dramatically increase our ability to execute tasks efficiently, but that does not guarantee that we are directing their efforts toward the right objectives.

I would argue that the fundamental unit of productivity in an agent-enabled organization should no longer be the individual employee or the individual agent. It should be the human-agent system responsible for producing a business outcome.

We need to measure the value of the outcome, the resources required to produce it, and the amount of human attention consumed along the way.

Otherwise, we risk creating organizations where machines are extraordinarily busy while people become overwhelmed managing their output.

The Junior Share of Technical Hires Fell to 9 Percent

There is another finding in Anthropic's research that deserves attention. The proportion of junior technical hires fell from 27% in early 2025 to 9% in 2026. The overall share of junior employees within technical departments also declined, from 21% to 14%.

This does not mean Anthropic stopped hiring technical employees. Its technical workforce continued growing in absolute terms. But the composition of hiring is changing, and some hiring managers explicitly reported recruiting fewer people because of AI.

Anthropic Figure 1  Decline of entry-level technical hiring. The junior share of hires, headcount, and open roles in technical and non-technical departments. Shares are distinct from absolute employment levels.

Source: Anthropic, October 16, 2026. Research report · Original figure

The company is also changing what it looks for in candidates. Hiring evaluations increasingly emphasize autonomy, judgment, and the ability to use AI creatively, while placing relatively less emphasis on narrow technical expertise.

I understand the logic. If AI agents can perform much of the routine coding and analytical work, companies may prefer employees who can operate independently, work across disciplines, and direct AI systems effectively.

But I keep coming back to one question: How do you become a senior engineer if you never spend time being a junior engineer?

A junior engineer learns much more than how to write code. They learn how systems behave, how to troubleshoot failures, how to recognize weak assumptions, and how to develop judgment through experience. The same applies to junior consultants, lawyers, analysts, and many other knowledge workers.

If agents increasingly perform the work through which people traditionally develop expertise, companies will need to rethink how they train the next generation.

Otherwise, we may create a strange situation where organizations become more efficient at producing work today while weakening their ability to develop experienced professionals tomorrow.

This is another reason I don't think agent management should be treated purely as a technology discipline. It also has to be a talent management discipline.

An effective agent-enabled organization must decide not only what work should be automated, but what experiences humans still need in order to develop the judgment required to supervise increasingly capable systems.

Hiring Managers Put More Weight on Judgment and Autonomy

What I find most valuable about Anthropic's research is that it moves the conversation beyond speculation.

We are no longer simply imagining a future where employees supervise AI agents. We can observe an organization where that transition is already taking place.

Employees are directing large amounts of machine-generated work. They are building systems to monitor other systems. They are spending more time evaluating outputs and resolving exceptions. Hiring managers are placing greater emphasis on judgment and autonomy. And the organization is confronting new questions about how to develop talent, allocate human attention, and measure productivity.

None of this means Anthropic's experience can be copied directly by every company. Anthropic develops frontier AI technology, employs people who are unusually comfortable working with AI, and operates in an environment where experimentation is part of the company's culture. A hospital, retailer, manufacturer, or financial institution will face very different constraints.

But that is exactly why I think we need a more systematic theory of agent management.

We should begin studying the relationship between agent autonomy and human supervision, the economics of human-agent productivity, the organizational structures that support effective delegation, and the ways companies can preserve human expertise as more execution is automated.

Perhaps the most important question is whether we can develop management systems that allow agents to become more autonomous without making humans less effective.

For more than a century, management theory has focused on coordinating human effort. The industrial revolution required us to rethink the organization of physical labor. The knowledge economy required us to rethink how we manage expertise and intellectual contribution.

Now we are entering a period where software can perform meaningful portions of knowledge work itself.

Anthropic's experience suggests that the next challenge will not simply be making agents more capable. It will be learning how to manage the work they perform, how to allocate authority responsibly, and how to ensure that the growing volume of machine execution translates into genuine organizational value.

And that may be where the next chapter of management theory begins.

Primary source: Nathan Wilmers, Regina Catipon, Maxim Massenkoff, and Peter McCrory, How Employment Is Changing at Anthropic, Anthropic, October 16, 2026.

The management interpretations and proposed research questions in this essay are my own. Anthropic's internal findings provide an early case study, not a universal model for every organization.