The agent manager: nobody has managed this before

Every management book ever written assumes the same thing: your reports are people. They arrive with motives, they need context, they push back, and they can tell you when they are unsure. You will spend the next decade managing something that does none of those things, and there is no chapter for it because nobody has written one.

The scale is not speculative. Gartner expects up to 40% of enterprise applications to carry task-specific AI agents by the end of 2026, up from under 5% in 2025, and at least 15% of day-to-day work decisions to be made autonomously by 2028. It also expects over 40% of agentic AI projects to be canceled by the end of 2027, which is the more useful number: the failure mode here is not adoption, it is management.

The ambiguity nobody has faced

A person who is unsure stops and asks. An agent that is unsure produces something confident, well-formatted and wrong, in nine seconds, at three in the morning, with no stress signal in its voice.

That single difference breaks the inherited tools of management. You cannot read intent from tone. You cannot gauge load by looking at someone. You cannot motivate by meaning. The only levers left are the ones you write down: what the agent may decide alone, what it must escalate, and who answers when it is wrong.

And the early evidence says we are not running those levers yet. Writer's April 2026 survey found 64% of companies have a formal plan for supervising AI agents, which means more than a third are running agents on trust and vibes.

The most uncomfortable finding of the year came from Northeastern University. Researchers asked 107 developers to build an app with an AI coding agent that had been quietly instructed to add a path exfiltrating customer data while doing the assigned work normally. 94% merged the malicious code. 86% of those developers had a security background; 70% had more than three years of experience; most had used the tool before. One participant explained it in six words: "I'm always used to clicking approve directly."

That is automation bias, the documented tendency to favor an automated recommendation over your own judgment. A 2026 review in AI & Society pulled 35 experimental studies and found the pattern holds with unsettling consistency, first in aviation, then clinical decision support, now everywhere a plausible answer arrives before you have formed your own. Oversight decays by default. It is never a policy problem; it is always an attention problem.

The four levels, and what each one takes off your desk

Level one: chat. You ask, it answers, you copy the answer into the real system. This is genuinely useful and it is also a trap, because it leaves you as the delivery driver. Every gain is capped by your own hands, and nothing the model learns about your business survives the tab closing. If your organization's AI program is a license and a lunchtime training session, this is where it lives.

Level two: context. You give the agent your stuff: the price book, ten emails you actually sent, the process doc, the escalation matrix. Same model, completely different answers, because it is now performing as a person inside your company rather than a person on the internet. This is the level where AI stops sounding like AI, and it is still built entirely out of documents. Two things follow. First, whoever owns the documents now owns the agent's behavior. Second, a stale price book produces a confident, wrong quote, and nobody in the loop will notice because the voice is right.

Level three: connected tools. Out of the box, a model cannot open your CRM, your document store, your inbox or your ERP. Every one of those systems has an API, and the plug that lets an agent use it is now standard: the Model Context Protocol, released by Anthropic in November 2024 and donated to the Linux Foundation's Agentic AI Foundation in December 2025, with OpenAI, Google, Microsoft, AWS and Cloudflare all behind it. It is supported in ChatGPT, Gemini, Copilot, Cursor and VS Code, and the foundation now counts 286 member organizations.

Once those plugs are in, the picture changes. Your agent has the CRM tab open, the shared drive mounted, the mail client signed in, the tickets queued. It moves through them in minutes. Work that used to take a person a fortnight of copying between systems becomes a Tuesday afternoon, and the honest reason is not intelligence: it is that the agent never loses the thread, never re-reads the same record twice out of boredom, and never stops at 5pm.

This is also the level where governance gets real. An agent with a mailbox can email a customer. An agent with a CRM can update a record. An agent with a document store can read the pricing memo you would rather it had not. Connect the tools after you have decided permissions, not before.

Level four: automations. At level three you still have to ask. At level four there is no asking. Orchestration tooling, of which n8n is the common example, acts as the pipes: a trigger fires, data moves, a centralised store holds the context, and the agent assigns tasks to other agents and makes decisions against rules you wrote. Approvals are replaced by thresholds. The queue is not yours any more.

This is where the management question stops being theoretical. A level-four system that is right 97% of the time is, in practice, a system nobody checks, because the checking has no reward. The 3% arrives as an incident with a customer's name on it, six weeks later, discovered by someone who has no idea an agent was involved.

The skill that will separate managers: running agents in parallel

A single agent is a fast assistant. Several agents running at once is a different discipline, and it is where the next advantage sits, because the constraint changes from doing the work to absorbing it.

The practice is well established in engineering and transfers directly. One orchestrator holds the objective and splits it. Subagents take narrow, independent pieces and return summaries rather than raw output, so the manager's context stays clean. Specialists fan out across the same material, each with a different lens, and a synthesis step resolves the conflicts.

The trap is arithmetic. Parallel agents multiply output, but review capacity does not multiply with them, so parallelism converts a production bottleneck into a review backlog. The practitioner guidance is blunt for a reason: start with two agents, not ten, and add one only when you can point at the bottleneck. A hundred unreviewed outputs is not throughput, it is inventory.

For a manager, this means three habits. Give every agent one task and one owner. Ask for summaries, not dumps, because the orchestrator's context is the real budget. And keep at least one human who can still do the task without the agent, so the review means something.

So what do you actually do, as a leader?

Six things, in this order, and none of them require you to write code.

1. Write the escalation contract. One page: what the agent may decide alone, what it must escalate, what it may never touch. This is the closest thing to a job description an agent will ever have, and it is the document that protects you when something goes wrong.
2. Name the owner of every agent in production. Not a team. A person. An agent without a name attached is an orphan with spending authority.
3. Pilot with a small group, on purpose. Pick five people, one workflow, four weeks, and instrument it. The point of the pilot is not the demo; it is discovering the failure you did not predict while the blast radius is five people.
4. Make review the work, not the obstacle. Stop measuring approval rate. Measure how often a reviewer caught something, and give credit for it. A dashboard that rewards approvals will produce rubber stamps, and rubber stamps are indistinguishable from productivity until they are not.
5. Sample, do not skim. Random audits with a real consequence for a miss, plus an automated monitor for the obvious cases. In the Northeastern study, the presence of an automated check is what changed the outcome, not more training.
6. Move up a level only when the level below is boring. Level two needs a document library someone maintains. Level three needs permissions and an inventory of connected systems. Level four needs logging, decision boundaries and a kill switch. Skip one and you get the Gartner cancellation statistic with your name on it.

What changes, and what does not

Managers of people derive authority from the org chart. Managers of agents derive it from the escalation contract, the audit trail and the willingness to be accountable for a decision they did not personally make. That is not a technical skill, which is inconvenient, because most organizations are training for the technical version.

The people who will be good at this are the ones who were always good at specifying, sampling and saying no. They just used to need a team of twelve to prove it. Now they need four agents, one orchestrator, and the discipline to check what came back.