Agent Orchestration Is an Organizational Problem

Organizational theory has something to teach us about AI agent orchestration. Multi-agent systems make delegation easy. They don't make coordination cheap.

Take span of control. The limit on direct reports was never about headcount — it's about how much coordination one manager can absorb before things slip. Routine work with stable interfaces lets a manager cover a lot of ground. Interdependent, uncertain work narrows that fast.

An orchestrator hits the same wall, dressed differently: can it split the work cleanly, judge what comes back, and untangle the results when they conflict? This is a known failure mode: agents launch subagents without distinct boundaries, produce overlapping or conflicting answers, and the reconciliation cost lands back on a human — the exact cost the system was meant to remove.

Handoffs fail for the same reason. A chat transcript is a weak way to pass work between specialised agents. The next agent needs a boundary object: task state, evidence, decisions already made — not a conversation it has to reinterpret from scratch.

Autonomy has to be designed, not assumed. Sign-off on every step and the system stalls. No sign-off anywhere and you get duplicate work, silent gaps, and results nobody checked.

Frameworks give you the plumbing — routing, memory, tool calls. They don't give you the arrangement: who has authority, what gets handed off, when a decision becomes an escalation. In a demo, skipping that is a shortcut. In an enterprise, it's a liability.

The moment an agent touches customer data, moves money, or changes production, authority is an access control decision, handoff is a data contract someone owns, and escalation is a check enforced in code, not a norm the agent might follow.

Most agent initiatives stop at the workflow for orchestration. The next bottleneck sits one layer up, and like autonomy, it won't resolve on its own. Span of control. Boundary objects. Coupling. Organizations solved these problems before agents existed, and a few people in organizational research are already making this argument for agentic AI.

Most orchestration effort stops at the workflow layer, and the next set of failures is already sitting one layer up.