AI agent orchestration: how handoffs work

AI agent orchestration is the coordination of several AI agents working on different parts of the same assignment. One agent might gather information, another check the findings, and a third prepare a finished document. Each receives work from the previous stage and passes its own result to the next. The arrangement sounds straightforward until something goes wrong. Who owns the finished result? Does the next agent receive the original instructions? Can one agent authorise another to take an external action?

For a founder or operations lead, good orchestration means dividing work without dividing accountability. A person owns the assignment, agents have defined responsibilities, and every handoff carries enough information for the next stage to proceed. The number of agents matters less than the clarity of those arrangements. Three agents with distinct jobs can be useful. Ten agents passing incomplete instructions between themselves can become another management problem.

The assignment needs one accountable owner

Consider a competitor research report. The assignment involves identifying competing businesses, checking their public information, comparing their offers and preparing a written assessment. One general-purpose agent might handle the entire task. Alternatively, a research agent could gather the information, a second agent could examine the evidence, and a writing agent could prepare the report.

The second arrangement separates the work, but it also creates a responsibility problem. The researcher owns the findings it produces. The checking agent owns its examination of those findings. The writing agent owns the document it prepares. None of them automatically becomes responsible for the business decision behind the assignment.

That responsibility stays with a human owner. They set the objective, decide what counts as an acceptable result and review the completed work. Agents are responsible for their assigned tasks, not for deciding whether the organisation should act on the outcome.

This distinction becomes useful when a report contains an unsupported claim. The owner should be able to trace that claim back to the research supplied to the writer. They can then see whether the researcher collected unreliable information, the checking agent missed it, or the writer changed its meaning. Without clear ownership at each stage, the only available explanation is that the agents produced a bad report.

A handoff must carry more than the last message

A handoff that says "Continue the research" gives the receiving agent too little to work with. It may not know which competitors matter, what the report is supposed to answer or which information has already been checked. Instructions that made sense at the beginning can disappear after several transfers.

Every receiving agent needs the relevant parts of the original brief and a clear account of the work completed so far. For a competitor report, that information should include:

• The original business question and the result the assignment must produce.
• The agreed scope, including which competitors and markets are included or excluded.
• Findings collected so far, with their sources and the dates they were checked.
• Unresolved questions, conflicting information and gaps in the evidence.
• The required deliverable, its intended reader and the deadline.

The receiving agent also needs to know what it may change. A writing agent can reorganise research into a readable report. It should not silently expand the scope, remove inconvenient findings or turn an uncertain statement into a confirmed fact.

Keeping the brief and previous results available also makes it easier to replace an agent. If a research task needs to be repeated, another agent should be able to pick up the documented work rather than reconstructing an assignment from an incomplete conversation. The work belongs to the delegation, not to the agent that happened to perform the previous step.

Give each agent a distinct job

Adding an agent should solve a specific problem. If two agents examine the same material and produce nearly identical results, separating their work may add little value. Every extra transfer creates another opportunity for information to disappear, instructions to change or work to be repeated.

Separate tasks when they require different skills, independent checking or different permissions. A support complaint, for example, might pass through three clearly defined roles:

The human owner reviews the proposed response and decides what happens next. If the complaint requires a refund, that is a separate decision. It should not become an automatic consequence of an agent preparing a sympathetic reply.

Notice that the agents do not need authority over one another. The workflow defines which task comes next and what information must accompany it. The research agent does not become a manager because the writing agent uses its findings. Each role has a limited responsibility, and the person overseeing the assignment retains control of the whole.

Put human approval at the point of action

Not every transfer needs human review. A research agent can pass its notes to a checking agent without someone approving the transfer. The checking agent can return corrections to the researcher. These are internal work steps, and an unnecessary approval at every stage would slow the assignment down.

An external action is different. Sending a message, paying an invoice, issuing a refund or publishing a document affects something beyond the workspace. Completing an earlier task does not give the next agent permission to take that action.

In Delagents, actions with an external effect wait in an approvals inbox. Each approval appears as a card showing the exact proposed action, the cost so far and a deadline. A person can approve it, change it or reject it. The agent can prepare the work, but the person decides whether the action goes ahead.

Eight categories of action always require human approval:

• Making first contact with anyone outside the workspace.
• Moving money or issuing refunds.
• Deleting or overwriting records the agent did not create.
• Changing roles or the agent's own policy.
• Connecting an integration or widening its access scopes.
• Accepting terms or signing agreements.
• Publishing material publicly.
• Taking action in another workspace.

Imagine a supplier research task. One agent identifies potential suppliers, an outreach writer prepares an introductory email, and another stage prepares the proposed send. The earlier work does not authorise contact. Approval belongs immediately before the first message leaves the workspace.

A workflow should identify these checkpoints before execution begins. Otherwise, an agent may interpret a completed handoff as permission to do something the original owner never authorised.

Keep a record of changes, costs and decisions

Assignments change while they are running. A source may turn out to be outdated, a manager may narrow the research scope, or an approval may be rejected. The next agent must receive the updated instructions rather than continuing with an earlier version of the brief.

The activity record should explain how the assignment reached its current state. Which tools did the agents call? What did a person approve? Was a proposed action changed or rejected? Did someone alter the policy? These details help the owner investigate mistakes and decide whether the work can continue.

An append-only activity log gives the team a record of tool calls, approvals and policy changes. It should be possible to examine what happened without relying on an agent's account of its own performance.

Costs need the same attention. If three agents contribute to one delegation, their combined work should remain within its hard budget cap. A handoff must not quietly create another allowance or hide what previous stages have spent. The owner needs the current cost before deciding whether further work is justified.

This matters when an assignment is paused or redirected. The owner can inspect the last completed stage, identify unresolved work and send the updated assignment to the appropriate agent. The next run starts with an accurate account of what has already happened.

Start with one assignment and two clear handoffs

A practical first orchestration plan does not need a large agent roster. Choose an assignment your team already understands, preferably one that repeats and has a deliverable you can inspect. A supplier comparison or a support response is easier to test than an entire department's workflow.

Write down the expected result and divide the work only where responsibilities genuinely differ. For every handoff, specify the information that must travel with the task, what the receiving agent must produce and what it cannot change. Mark external actions that need a person's approval.

1. Choose one recurring assignment with a clearly defined result.
2. Write the brief, including scope, evidence requirements, deadline and budget.
3. Assign agents to separate stages that need different responsibilities.
4. Specify what each agent must receive and return at every handoff.
5. Identify the decisions and external actions that require human approval.
6. Run the workflow and inspect where information disappeared, costs accumulated or work had to be repeated.

For teams putting together their first roster, the Delagents agent roster and builder provides access to agent roles and templates for research, support triage, bookkeeping, outreach, meeting preparation, inbox triage, content repurposing and recruiting screening. Delagents is in preview. Its working model gives an assignment to an agent through a human-written brief, stops at checkpoints for sign-off and delivers completed work as deliverables.

Take an assignment your team completed last week and map its first two handoffs. Write down what the first agent would need to receive, what it should return and which instructions the next agent must preserve. Then identify the exact point where a person takes responsibility for the next decision. If those boundaries are difficult to explain, keep the work with one agent until they are clear.