AI agents need org charts, not pipelines

Here’s what nobody in the multi-agent AI conversation is saying: the hard part isn’t the technology. It’s the org design.

Every major agent framework — CrewAI, LangGraph, AutoGen, OpenAI’s Swarm — organizes agents the same way: around tasks. You define a job. You spin up agents. They execute. They disappear. The framework treats the problem as a pipeline: inputs go in, outputs come out, agents are the pipes.

This works fine if your work is a pipeline. Most work isn’t.

I spent twenty years running consulting engagements at Fortune 500 companies — Disney, Delta, Home Depot, Wells Fargo. Not writing code. Managing teams of specialists who had to produce coordinated output under constraints. What I learned is what the AI agent community is about to discover the hard way.

Agencies don’t organize around projects. They organize around competencies.

This is the thing everyone outside the industry gets wrong. They assume agencies assign dedicated teams to each client. Some do, for the biggest accounts. That’s the expensive, unscalable model. The agencies that work organize by what people know how to do: strategy, research, design, content, QA. Each competency is a team. Each team serves every client.

The client engagement gets a project manager. The PM is the only person bound to the client. Everyone else rotates. The strategy team works on Delta this week, Disney next week. The designer handles three clients simultaneously. The competency is persistent. The assignment is temporary.

One great strategist serves ten clients. One great reviewer serves eight. That’s how you scale without scaling headcount linearly.

The brief is the interface. When the strategist hands work to the writer, she doesn’t say “write something good.” She hands over a document: here’s the audience, here’s the objective, here’s the constraint, here’s what done looks like. The quality of the output is almost entirely determined by the quality of the brief. And review is adversarial, not ceremonial — the reviewer’s job is to break the work, not admire it.

Now look at your AI agent stack.

If you’re like most people, you have a dedicated agent per project. One agent handles your newsletter. Another handles your research workflow. A third handles your CRM. Each one knows its project intimately. Each one knows nothing else.

That’s not an agency. That’s a freelancer for each client — the expensive, unscalable model. The one the smart agencies abandoned. A single agent with unlimited scope is the AI equivalent of a junior consultant told to “just handle everything.” It can do anything. It has no idea what it should do next.

Three paradigms. Most systems are stuck in the first one.

The ephemeral swarm is what most frameworks ship. Spin up, execute, tear down. No memory, no institutional knowledge. Works for isolated tasks. Falls apart for anything requiring continuity.

The long-lived project is better — persistent context about a single domain. But it creates silos. Your QA agent for one product has never seen your other products. Your reviewer can’t notice when the same mistake shows up across three different workstreams.

The long-lived role is what nobody has tried yet. A QA agent that rotates across all your work. A researcher that builds expertise in research — not in one domain — and brings that to every assignment. The competency persists. The assignment rotates. Same as the agency model. Same benefits.

Here’s the inversion most people are missing. Right now, almost everyone treats the project as context and the role as a skill — a prompt label you paste in at the top. “You are a QA engineer.” The project is what persists. The role is what’s cheap.

Flip it. Make the role the context — the accumulated expertise, judgment, and identity that persists across assignments. Make the project the skill — a brief, handed in, executed, handed back. The role is the expensive thing to build. It should persist. The project assignment is the cheap thing to hand over. It should be ephemeral.

Cheaper, because you’re not reloading an entire domain into every agent on every run. Better, because a role-context agent brings accumulated judgment to each new assignment — a QA agent that has broken things across twenty projects knows failure modes a project-bound agent never sees. Faster, because the brief is a fraction of the cost of rebuilding project context from scratch.

Here’s where it gets interesting for non-developer work.

In a software project, the codebase is the context. You need an agent that knows the repository intimately: the architecture decisions, the naming conventions, where the bodies are buried. Losing that between runs is expensive. The project-per-session model makes sense.

In non-developer work — content strategy, marketing, operations, consulting — the domain knowledge is portable. A writer who just drafted a newsletter can draft a course email five minutes later. A researcher doesn’t care whether she’s analyzing competitors for a SaaS product or pricing for a workshop. The function is the persistent context, not the project.

For non-developer work, you don’t need agents that know your projects deeply. You need agents that know their jobs deeply — and a manager that knows what to assign them.

This is what every Fortune 500 company is about to get wrong. They’re building dedicated AI agents for each business unit, each workflow, each product line. Project-per-agent, all the way down. The expensive, unscalable model.

The directional problem with the whole conversation

Every consulting firm, every AI platform, every voice in this space right now is asking the same question: how do AI agents change your organization? McKinsey’s “Agentic Organization.” BCG’s agent frameworks. Deloitte’s role-specific deployments. All of them going in one direction — from agents to org impact.

Nobody is going the other way.

Nobody is asking what organizational theory tells us about how to design agent systems in the first place. What we already know from management science, from thirty years of running agencies, from the basic mechanics of specialization and coordination — and could apply directly, right now.

The answer was settled decades ago. Organize by competency, not client. Make the brief the interface. Use adversarial review. Let the manager manage flow, not content.

These aren’t AI research insights. They’re organizational principles that predate the internet. The agent community is going to rediscover all of them, slowly and painfully, through expensive trial and error.

Or you could just talk to someone who ran an agency.