Andrew Chen put it cleanly on X: in a world of agents, the PM role fractures into two distinct jobs — one that organizes humans, and one that organizes agents.

Both are still chasing the same goal (ship the right product to customers), but the skills, the tools, and the mindset couldn't be more different.

PM Type 1: Human Organizer

This is the traditional PM job, still very much real.

Stakeholder alignment. Design reviews. Engineering standups. Prioritization meetings. Roadmap decks. Negotiating with four different teams about what gets built in Q3.

The soft skills here — reading the room, building consensus, translating between business and engineering, managing up — are fundamentally human. Agents make some supporting work faster (writing specs, summarizing threads, drafting updates), but the core job is human coordination.

PM Type 2: Agent Organizer

This is the new job that barely existed two years ago.

Instead of wrangling engineers and designers, you're wrangling agent pipelines. Your job is: prompts, evaluations, workflows, and failure modes.

What does this actually look like in practice?

  • Writing agent specs: instead of a PRD for humans, you're writing instructions for an agent that need to be precise enough to produce consistent output
  • Building eval systems: you can't just "review" agent output manually at scale — you need automated quality checks
  • Debugging pipelines: when an agent workflow breaks, the debugging process looks nothing like traditional software debugging
  • Prompt engineering: which is increasingly becoming a core competency, not a side skill

The Skills Overlap Is Smaller Than You Think

Traditional PM skills transfer somewhat. Systems thinking helps. Prioritization helps. Writing clearly helps.

But a lot of what makes a great human-organizer PM — political awareness, relationship building, emotional intelligence in high-stakes meetings — is irrelevant when your "team" is a set of agent pipelines.

And a lot of what makes a great agent-organizer PM — understanding token limits, knowing when an agent needs more context vs. less, writing eval rubrics — is completely foreign to traditional PM training.

What This Means for Solo Founders

If you're a one-person company, you're already doing both jobs.

You're organizing yourself (the human) and increasingly organizing a set of agents that do work in parallel. The question is whether you're doing the agent-organizing job deliberately — with proper evals, with reliable prompts, with Skills that actually work — or just hoping for the best.

The founders who treat their agent stack like a team they manage — with the same rigor they'd apply to managing humans — will build faster and more reliably than those who just "use AI" ad hoc.

The job has changed. The question is whether your working style has changed with it.