An AI agent handoff template that preserves the work

An agent handoff transfers a bounded piece of work with enough context for the next person or system to continue. It should state what is needed, what already happened and which authority applies. Sending an entire conversation is often less useful than a concise brief with links to evidence.

Make the recipient's job explicit

Imagine a research assistant passing a supplier comparison to a manager. The manager needs the decision criteria, the options, unresolved questions and links to the quotes. They do not need every search query or a vague message saying the research is done.

The handoff should also distinguish review from action. Asking a manager to evaluate a recommendation does not authorize the assistant to place an order afterward.

Copy the handoff structure

• Goal: the business result and the specific decision now needed.
• Current state: completed work, artifact locations and verified results.
• Open issues: uncertainty, missing information and known constraints.
• Authority: actions already approved and actions still requiring a decision.
• Next step: one concrete action with an accountable owner.

Include dates when facts may change. If a price or availability check is old, label it instead of allowing a fresh handoff message to imply fresh evidence.

Verify the handoff reached a usable place

A local file path is not a useful delivery if the recipient cannot open it. Use the agreed shared system and confirm the intended access. For machine-to-machine work, confirm that the receiving system accepted the payload and can parse the required fields.

Acceptance is not the same as task completion. Preserve separate states for handed off, acknowledged and resolved when the process needs them.

Close the loop without creating noise

Agree on where the result will be recorded and when an exception needs escalation. Do not require repeated status messages when nothing changed. A clear shared record gives the next owner continuity while reducing the burden of reconstructing a conversation from several channels.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.