MyDot Consultancy

Featured image: AI-generated conceptual illustration. Sources checked 26 October 2026.

“Research the supplier and send me what you find” is a weak delegation instruction for a person and an even weaker contract between software agents. It leaves the scope, evidence standard, permissions and completion criteria unclear. A useful delegation packet specifies the work and the form of the result before the worker begins.

In a multi-agent system, the packet is part of the application design. It determines what context crosses the boundary, what the worker may do and what the coordinator can safely conclude from the response. Treat it with the same care as an API contract.

Define the assignment in business terms

Imagine a hypothetical procurement comparison with separate workers reviewing product capabilities, support terms and implementation requirements. Each worker needs a bounded question. “Assess support terms for the named products using current official documentation” is clearer than “be the support expert.”

Include exclusions. The support worker should not contact vendors, negotiate terms or infer private contract provisions. It can identify what public documentation says and what requires confirmation. A role name does not establish those boundaries by itself.

Anthropic’s multi-agent research case study describes a coordinator working with separate research agents. The practical design lesson is to make the division of work explicit. The case study’s results should not be generalized into a promise for an unrelated workflow.

Specify the input contract

The worker should receive the objective, relevant entities, time horizon, available sources and applicable constraints. Pass stable identifiers where possible. A product name without an edition or region may be insufficient for a reliable comparison, and a customer name may be ambiguous.

Provide only context needed for the subtask. A worker checking public support documentation does not need confidential pricing negotiations or employee personal information. Reducing unnecessary context can improve clarity and limit exposure, but do not omit a constraint that changes the meaning of the task.

If the assignment depends on another worker’s result, make that dependency visible. A technical-requirements review may need the selected deployment option. Launching it before that option is known can create contradictory assumptions that the coordinator later mistakes for factual disagreement.

Specify the output contract

Ask for findings with source references, relevant qualifications and unresolved questions. Distinguish a source-backed fact from an interpretation. Require a clear status such as complete, partial or blocked, together with the reason. A polished narrative alone makes it difficult to detect missing coverage.

For the support comparison, a worker might return a documented contact channel, the relevant plan scope, the date checked and a note that response-time commitments were not publicly available. The absence of a commitment is a finding about available evidence, not proof that no private commitment exists.

The OpenAI Agents SDK handoff documentation covers handoff inputs and filtering. These mechanisms can help structure a transfer, but they do not supply the business contract automatically. The application must decide what the receiving agent needs and what it is allowed to conclude.

Assign ownership of shared state

Workers that only return findings are easier to coordinate than workers that all edit a shared report. If shared editing is necessary, assign sections or use proposed changes that a single owner merges. Record which version each worker used so a stale update does not overwrite newer work.

External actions need even clearer ownership. One worker should not send a message while another is still revising its content. A coordinator should not assume that a worker’s statement “ready to send” means the message was approved or sent. Keep preparation, authorization and execution distinct.

The SDK orchestration guide describes different coordination patterns. Choose a pattern that matches the ownership model. A deterministic merge or approval stage can be preferable to giving several agents discretion over the same destination.

Preserve disagreement instead of smoothing it away

Two workers may report different facts because they read different versions, regions or product editions. Require the coordinator to compare scope and sources before synthesizing. A compromise sentence can be less accurate than either original finding if it hides the conditions that made them differ.

If one worker reports a feature and another cannot verify it, the coordinator should not count that as a simple majority vote. Inspect the evidence. The right result may be a qualified statement or an explicit gap requiring vendor confirmation.

A worker’s confident tone is not an evidence standard. The coordinator should reject unsupported claims or request clarification when the output contract is not met. Delegation moves work; it does not remove responsibility for the final result.

Give failure a useful shape

• Blocked source: identify the inaccessible source and what conclusion cannot be reached.
• Partial coverage: list completed entities and remaining entities.
• Conflicting evidence: return the conflicting claims with their source scopes.
• Budget reached: preserve verified findings and state what remains unexamined.
• Denied action: record the boundary without seeking an alternative route around it.
• Canceled task: stop new work and report any already committed external effects.

The coordinator should have a policy for each status. It may proceed with a clearly labeled partial report, ask the user for missing information or stop because the absent evidence is essential. Do not silently replace a failed worker with speculation.

Test the handoff itself

Use a controlled assignment with ambiguous names, a missing source and conflicting documents. Inspect whether the worker asks for clarification, preserves limitations and returns usable references. Then test whether the coordinator retains those limitations in the final artifact.

Also test context filtering and permissions. A narrow worker should not inherit unrelated confidential material or tools simply because the parent has access. Review what actually crosses the handoff boundary, not only what the assignment text says should cross it.

Make delegation reviewable

Keep the assignment, returned result and coordinator decision linked in the task record. An operator should be able to explain why a subtask existed, what it established and how its findings affected the final output. This makes disagreements and repeated work diagnosable.

A strong delegation contract is concise but complete: objective, scope, authority, evidence and result. With those elements in place, multiple agents can contribute useful work without losing the qualifications and accountability that a business needs from the final answer.

Sources and further reading

Anthropic: How we built our multi-agent research system. First-party coordinator and research-worker architecture; published June 2025.

OpenAI Agents SDK: Handoffs. Handoff inputs and filtering mechanisms; checked October 2026.

OpenAI Agents SDK: Agent orchestration. Coordination pattern documentation; checked October 2026.