The New Management

Here is a paradox worth sitting with. We built AI agents to reduce our dependence on coordination. And the skill they demand most is coordination.

AI agents are extraordinarily fast, moderately reliable, and have absolutely no sense of direction. They will build the wrong thing with the same enthusiasm as the right thing. They will not raise a hand to say the spec is contradictory. They will not notice that two parallel tasks are producing incompatible outputs. They will work at three in the morning without complaint, and by sunrise we may have six features that do not fit together.

This is not a tools problem. This is a management problem. And the people best prepared to solve it are not the best coders. They are the best managers.

The Paradox of More

Google recently demonstrated something that should give the entire agentic AI industry pause: adding more agents to a system can make it worse. Not marginally worse. Categorically worse. The multi-agent architectures that look elegant on a whiteboard, where agents communicate with each other, negotiate task allocation, and coordinate in real time, produced worse outcomes than simpler configurations.

The frameworks that recommend inter-agent communication are, in many cases, simply wrong. Gartner now predicts that 40% of agentic AI projects will be cancelled by 2027. Not because the technology failed. Because the orchestration did. The agents worked fine individually. Nobody managed the ensemble.

The systems that actually scale look nothing like what the thought leadership recommends. They look boring. One human, clear specifications for each agent, strict isolation between tasks, and a coordination layer that lives entirely in the manager’s head. Not a swarm. A team with a lead.

We have seen this before, of course. Every generation of management theory rediscovers that adding people to a late project makes it later. Adding communication channels between teams increases overhead faster than it increases throughput. The same laws apply to agents, only faster, because agents do not have the social intelligence to self-correct when coordination breaks down.

The Skill That Was Already There

Large enterprise AI rollouts consistently reveal the same pattern. The majority who stop using AI tools are not the least technical. They are the ones who lack what researchers have started calling “management instincts” for AI.

That phrase is worth pausing on. Management instincts. Not prompting techniques. Not technical fluency. The ability to delegate clearly, check work at the right intervals, know when to intervene and when to let the work run, and maintain a mental model of how the pieces fit together.

Teams using AI coding tools consistently report the same inversion: senior engineers accept more AI-generated output than juniors do. Not because seniors are less careful. Because they are better at specifying the task, evaluating the output against their mental model, and correcting course before the error compounds. They have, in other words, management experience. Years of coordinating work, reviewing deliverables, and knowing when “good enough” is actually good enough.

The irony is beautiful. The skill that was supposed to become obsolete in the age of AI (middle management, coordination, the soft art of running a team) turns out to be the skill the age of AI needs most.

What Orchestration Looks Like

Steve Yaggi, a single engineer, orchestrates 20 to 30 AI agents simultaneously. One person, massive output. This sounds like a productivity story, but it is actually a management story. What Yaggi is doing is not coding. He is running a team. He is decomposing work (the previous skill in this series), assigning it, monitoring progress, catching conflicts early, and integrating the outputs into a coherent whole.

The practical shape of this skill has a few consistent patterns.

Isolation is the first principle. Agents that talk to each other produce emergent confusion. Agents that receive clear, independent specifications and return completed work to a human coordinator produce results. The instinct is to let the agents collaborate. The reality is that collaboration requires social intelligence that agents do not have. We provide the social intelligence. They provide the speed.

The second pattern is rhythm. Good orchestration is not continuous attention. It is periodic checkpoints. Let three agents run for twenty minutes, review the outputs, catch the drift, correct the specs, launch the next round. This is not micromanagement. It is the same cadence any good manager uses with a high-performing team: enough oversight to catch problems early, enough autonomy to let the work flow.

The third pattern is knowing what to check. We cannot review every line an agent produces, and we should not try. The skill is knowing which outputs are high-risk (architectural decisions, data model changes, anything that other work depends on) and which are low-risk (formatting, boilerplate, repetitive transformations). Experienced managers already think this way. They do not review every email their team sends. They review the ones that matter.

The fourth pattern is managing context. Research suggests that results degrade noticeably once an agent’s context window is about 40% utilized. Skilled orchestrators use separate context windows for separate concerns: one for research, one for design, one for implementation. They persist important information in static artifacts (documents, not conversation history) so that each new context window starts clean. This is not a technical optimization. It is the equivalent of not letting one meeting bleed into another.

The fifth pattern, perhaps the most important, is admitting when the coordination has broken down. When three agents have produced work that does not integrate cleanly, the answer is almost never to add a fourth agent to reconcile them. The answer is to step back, simplify the decomposition, and try again with clearer boundaries. Good managers know when to reorganize the work, not just the workers.

This Is Not New Work

Here is the permission this article is meant to give.

If we have spent years managing people, coordinating projects, running standups, reviewing work, aligning teams toward a shared goal, and wondering whether those skills still matter in a world that seems to value only technical fluency: they matter more now than they did last year. They matter differently, but they matter more.

The conversation around AI has been dominated by technical skill for two years now. Learn to prompt. Learn to code. Learn the new frameworks. And those things are real. But the bottleneck we keep finding is not technical. It is organizational. The teams that get the most from AI are not the ones with the best engineers. They are the ones with the best coordination. Someone who knows how to scope work, distribute it, review it, and integrate it. Someone who has been doing exactly that, with human teams, for years.

Orchestration is not a new skill. It is an old skill applied to a new medium. The medium is faster, more literal, and less forgiving of ambiguity. But the core capability, the ability to coordinate parallel work streams, maintain quality across a distributed effort, and know when to intervene versus when to trust the process, that is management. It always was.

The K-shaped economy is not just splitting along technical lines. It is splitting along managerial lines. The people who rise are not only the ones who can code. They are the ones who can run a team, whether that team is made of people, agents, or both. And the ones who recognize this early have a compounding advantage, because every month the agents get faster and the need for good orchestration grows with them.

The agents do not need a better algorithm. They need a better manager. And the best managers, it turns out, have been practicing this skill for decades. They just did not know they were training for this.

This is part of “The Hitchhiker’s Guide to the K-Shaped Economy.” Previous: “Think in Pieces” on decomposition. Next: on intent.