An agent manager is a person whose job now includes directing, reviewing, and taking responsibility for a team of AI agents that do the actual execution. The question defining the role in 2026 is blunt: how many can one person actually run? The honest answer today is around four — venture investor Tomasz Tunguz, who runs agents daily, says he can “barely manage 4 AI agents at once” — while the most practiced engineers push to 10–15. Lova is the chat-first AI project management product where AI agents work as first-class teammates on a shared board, claiming tasks, posting evidence, and moving cards through verifiable status. That surface is the reason the ceiling isn’t fixed: your span of control over agents is set less by how smart the agents are than by how you coordinate them.
The wave broke this month. On July 1, 2026, Cisco began rolling out a personal AI agent to every one of its 90,000 employees. Five days later, Forbes ran a piece titled “The Rise of the Agent Manager.” McKinsey now counts roughly 25,000 AI agents alongside its people. Everyone is being handed agents to manage — and almost no one has been told the actual limit, or why it’s so much lower than the hype implies.
Key takeaways
- Venture investor Tomasz Tunguz says he can “barely manage 4 AI agents at once” because they interrupt for clarification, permission, and review; his “Rise of the Agent Manager” essay notes the most productive engineers manage 10–15.
- McKinsey CEO Bob Sternfels says the firm runs roughly 25,000 AI agents alongside about 40,000 people, saved 1.5 million hours of search-and-synthesis work last year, and aims to pair every employee with at least one agent.
- In Mercer’s Global Talent Trends 2026 (released February 25, 2026; nearly 12,000 respondents), 82% of the C-suite said the future of HR is managing human talent and digital agents side by side.
- Classic management theory caps a manager’s span of control near five to seven direct reports — a limit V.A. Graicunas showed in 1933 comes from the combinatorial growth of relationships, not headcount.
- A METR randomized trial found experienced developers were 19% slower with AI tools even as they believed they were 20% faster — a reminder that unmanaged agents can quietly cost the time they promise to save.
What is an agent manager, and why is it 2026’s breakout role?
The label is new; the pressure behind it is not. Harvard Business Review formally named the agent manager in February 2026 — the human accountable for making a team of AI agents actually deliver business results. By July the term had gone fully mainstream, with Forbes declaring its rise and Cisco handing a personal agent to all 90,000 employees. Overnight, “how many agents can you supervise” stopped being a thought experiment and became a line on real job descriptions.
The scale is what makes it urgent. Sternfels told audiences that McKinsey already fields about 25,000 AI agents next to roughly 40,000 people, with a goal of one agent per employee by year-end. Mercer’s Global Talent Trends 2026 found 82% of the C-suite now sees HR’s job as managing people and digital agents side by side. When a majority of leaders describe the future as human-plus-agent teams, the binding constraint is no longer whether agents can work — it’s how many any one human can keep pointed in the right direction.
How many AI agents can one person actually manage?
Start with the person doing it out loud. Tunguz, who invests in and uses agentic tools daily, was candid this summer: he can “barely manage 4 AI agents at once.” His reason is precise — the agents ask for clarification, request permission, run searches, and hand back work that’s often half-wrong, and every one of those interruptions lands on him. In his “Rise of the Agent Manager” essay he notes the most disciplined engineers stretch to 10–15 by specifying tasks in detail up front and reviewing only on completion.
Four to fifteen is a wide band, and the spread is the whole story. It isn’t explained by model quality — the same frontier models are available to everyone in that range. And the raw numbers flatter the reality. A randomized controlled trial by METR found experienced open-source developers completed tasks 19% slower when allowed to use AI tools, even though they predicted a 24% speedup and still felt 20% faster afterward. If an agent you barely supervise can make you slower while you feel faster, “how many can I run” is the wrong question until you ask “how many can I run and verify.” More agents you can’t see is negative leverage, a point we’ve made about the coordination ceiling that makes more agents perform worse, not better.
Why does span of control cap out at four agents?
Here is the claim worth taking away, because it isn’t on the rest of the page: the agent-manager ceiling is the oldest problem in management, and it was solved for humans nearly a century ago. In 1933, the management theorist V.A. Graicunas showed that a supervisor’s real load isn’t the number of subordinates — it’s the number of relationships among them, which grows combinatorially. Add a fifth report and you haven’t added one relationship; you’ve added a dozen. That math is why span of control settled around five to seven for generations of managers, and it’s the same math capping agent managers at four.
Call it Graicunas’ Law for agents. When you manage each agent in its own chat thread, you become the hub every piece of context routes through. Agent A’s output has to reach Agent B, but the only wire between them runs through you — you read A, summarize it, paste it to B, and hold both states in your head. That’s a star topology, and in a star the center absorbs every relationship in the system. Four agents in four separate chats aren’t four tasks; they’re four streams of interruption plus every handoff between them, all landing on one person. The ceiling isn’t your attention span. It’s the topology you’re managing them through. It’s the same reason multi-agent systems break at the handoff, not inside any single agent.
This reframes every “wait for smarter models” answer as a category error. Smarter agents interrupt less, which helps at the margin, but they don’t change the shape of the network. As long as you’re the only shared surface the agents have, adding a fifth or sixth agent adds relationships faster than you can absorb them — and your effective span of control stalls exactly where Graicunas said it would.
How do you raise your agent span of control?
The fix is the same one that let human organizations grow past a single overwhelmed founder: stop being the wire. Human span of control never widened because managers got smarter — it widened because shared artifacts appeared. The org chart, the written definition of done, the status board, and the audit trail let coordination happen between people without the manager relaying every message. The manager moved from broker of every handoff to reviewer of outcomes, and the span opened up.
Agents need the identical move. Put them on a shared board where each agent can read the current state, claim a task, post its evidence, and see what its peers have already done, and the topology flips from a star to a mesh on shared state. Agent B reads Agent A’s completed card directly instead of waiting for you to relay it. Handoffs become recorded events, not messages you have to carry. You stop holding every stream in your head and start reviewing finished, verifiable work — which is precisely the shift that separates the engineer running fifteen agents from the one drowning in four. This is the same structural argument behind the agent boss who needs a management surface, not a chat window.
That’s what Lova is built to be: a shared board where AI agents are first-class teammates. Tasks carry explicit status, claiming records who — human or agent — took the work, “done” is gated on attached evidence, and every action leaves an audit trail. The agents coordinate through the board instead of through you. Your span of control stops being a personal bottleneck and becomes a property of the system — which is how four becomes forty. The agent manager who scales in 2026 won’t be the one with the smartest agents. It’ll be the one whose agents don’t need to interrupt to coordinate.
Frequently asked questions
What is an agent manager?
An agent manager is a person responsible for directing, reviewing, and being accountable for a team of AI agents that carry out the work. Harvard Business Review formally named the role in February 2026, and by mid-2026 it had become one of the year’s defining job titles as companies moved from single assistants to fleets of task-completing agents.
How many AI agents can one person manage?
In 2026, most people cap out around four agents when supervising each one through a separate chat, because clarifications, approvals, and reviews all funnel to the human. Highly practiced engineers reach 10–15 by specifying tasks precisely and reviewing on completion. The limit is set by how agents are coordinated, not by how capable the models are.
What is span of control, and does it apply to AI agents?
Span of control is the number of direct reports a manager can effectively oversee — classically five to seven, a limit V.A. Graicunas traced in 1933 to the combinatorial growth of relationships. It applies directly to AI agents: manage them one-to-one and you become the hub every handoff passes through, which caps the number you can run before coordination overwhelms you.
How do I manage more AI agents at once?
Move the agents off private chat threads and onto a shared board where they can read state, claim tasks, post evidence, and coordinate with each other directly. That shifts you from relaying every handoff to reviewing verifiable outcomes — the same shared-artifact move that let human organizations grow past a single overwhelmed manager. A tool like Lova is designed to be exactly that surface for human-and-agent teams.