Historically, the main constraint on software teams has been their ability to ship software. Teams could start with extremely ambitious missions that translated into clear visions of the future. They’d set roadmaps and goals managed by layers of managers, and then individuals on the team would translate those ideas into something tangible for customers.

But in the age of AI agents, I don’t believe this is the constraint anymore.

I’ve been using Claude Code’s new subagents feature quite a bit lately. For me, it feels like the ChatGPT moment for agents: creating a multi-agent system that can operate on complex tasks autonomously no longer requires tremendous amounts of custom software. You can literally create one with a few markdown files.

This shift is what Andrew Ng calls the “Product Management Bottleneck”. As he points out, “Because highly agentic coding accelerates the writing of software to a given product specification, deciding what to build is the new bottleneck, especially in early-stage projects.”

In this world where an agent can represent the persona or expertise of a specific type of employee, so much of the constraint becomes figuring out how to architect the optimal team for your project.

Instead of a manager needing to hire a team, onboard them, and hope they’ll be productive within the next few months, they can literally specify the shape of the team they need to deliver on an idea. The scope of management in this era is to:

  • Define the shape of that team
  • Create clear specifications and project plans
  • Have the expertise to observe and intervene as work progresses
  • Make sure everything heads in the right direction

What we’ve classically called the science of management becomes the bottleneck. Being able to translate your idea into a specification, that specification into a team shape, define that team’s work, manage the delivery, and ultimately execute: all of this can now be done by one enterprising individual with an orchestration layer for agents like Claude Code.

As Ng notes in his piece on AI Product Managers, “Given a clear specification for what to build, AI is making the building itself much faster and cheaper. This will significantly increase demand for people who can come up with clear specs for valuable things to build.”

In the agent world, product management and people management are becoming the same discipline. When you’re orchestrating agents, you’re simultaneously:

  • Deciding what to build (traditional PM work)
  • Deciding who builds it (traditional people management)
  • Designing how they work together (organizational design)

The traditional Engineer:PM ratio of 6:1 that Ng mentions is already obsolete when your “engineers” are agents that you can spin up instantly.

Do we need new Peter Drucker-style management books written specifically for managing agents? Or does all our prior art in behavioral science and management theory apply to this new world?

Does Andy Grove’s definition of high output management, where a manager’s output is the output of their organization plus the neighboring organizations they influence, map to what’s required now?

Some management challenges I’m already facing with agents:

  • How do you know whether an agent is succeeding when you’re the one who programmed it?
  • What does it mean to “fire” an agent if it’s underperforming?
  • How do you measure performance across different agent personas?
  • When is it time to “hire” new agents or scale up certain agent types?

All the classical tasks that managers learn hands-on by working with humans will likely be greatly accelerated and run in more simulated environments.

Agent management is a new unlock for team productivity, but it might also be the training ground for our next generation of leaders to learn how to manage humans.

Of course, subservience, adherence, and work ethic manifest very differently in human versus agent workers. But the core skills translate: clarity of communication, systems thinking, performance measurement, team composition.

There’s still tremendous experimentation and research to be done. But whatever the result, I believe that learning how to be a manager in the age of AI agents is one of the most important skills to develop. As a current people manager, it’s been eye-opening. But I think it’s one of the biggest opportunities for people earlier in their careers to differentiate themselves as they enter the workforce.

Ng writes that “PMs with high user empathy can make decisions by gut and get them right a lot of the time,” and that matches my experience. In the agent world, that translates to having high “agent empathy”: understanding what your artificial team members can and can’t do, and making rapid decisions about how to deploy them.

We’re not trying to replace human creativity or judgment, merely amplify it. When I can prototype an idea with a team of agents in hours instead of months, I can test more hypotheses and explore more possibilities.

The future is humans orchestrating agents to achieve things we never could before. That’s a future worth managing toward.