Skip to content
CARE

AGENTIC CARE · WHITEPAPER

When AI Agents Manage Other AI Agents

Supervision, delegation, and authority in coordinated systems

A supervisor agent can dispatch work. It cannot grant itself, or its workers, an authority the organization never gave.

A supervisor agent also needs supervision

Coordination creates dependencies between agents, not merely independent tasks.
Coordination creates dependencies between agents, not a stack of independent tasks.

An orchestrator dispatches tasks, workers call tools, and another agent scores the output. The business sees one service. The operation is a chain of delegated capabilities. Adding the orchestrator feels like adding a manager. It adds another actor that can request actions it was never allowed to authorize.

An agent hierarchy does not create a human accountability hierarchy. The organization has to write one, at the boundary where work leaves one agent and becomes someone else’s action.

Specialist agents hand work across boundaries

One agent coordinates specialists. It asks a second to summarize customer records and a third to send an action request. No single agent intended the combined outcome. Each step was inside its own instructions.

The failure mode is specific. The coordinator treats “send” as a task it may assign. The sender treats the assignment as authority. The customer receives a commitment nobody’s owner accepted. A second model that reviews the first can share the same missing fact and approve a fluent explanation.

In a demonstrated accounts-payable handoff, the outputs are named and the consumers are named: the ERP sink, the tax-adviser sync, and a human review. The map makes the boundary visible. It does not, by itself, decide who may widen it.

Write the delegation before you add the hierarchy

CARE process map highlighting handoff outputs to ERP AP Sink, Steuerberater Sync, and human review consumers for Accounts payable intake.
Observed CARE process map for accounts-payable intake: handoff outputs to ERP AP Sink and Steuerberater Sync, plus human review. The capture shows the declared consumers, not a completed transfer of authority.

The owner of the business result — the person who would answer for the message, the record, or the payment — writes the delegation for one coordinated service before another supervisor is added.

  1. Name who may delegate, and the actions they may request. “Coordinate the case” is not a list.
  2. Name what cannot be relaxed by the coordinator: prohibited acts, data that must not move, and any external commitment.
  3. Name who accepts the result that leaves the company. A worker agent does not inherit that acceptance by being asked.
  4. If a reviewer agent is in the chain, give it evidence the actor did not write: the system record, or a human decision. A paraphrase of the actor is not a review.
Build the next agent only inside that written delegation. A coordinator may propose a wider scope. It may not grant it.

Human authority sits on the external result

You can run a network of agents at the speed of the work, because the authority question was answered once, at the boundary, and is reused.

The result you can show today is one page for one service: delegator, prohibited relaxations, and the human who accepts the external effect. Where a cell is blank, that action does not run. Repeat the page when a new worker is added. Oversight stays practical because you are not re-deciding every handoff, and you are not pretending the transcript of the swarm is the management record.

Take action now

  1. Tell your agent to register itself in CARE: https://ai-managementsystem.com/under-care.
  2. Let the agent establish the management baseline.
  3. Use the outcomes.

Images and text generated with the assistance of ChatGPT by OpenAI.

Download