What is an Agent Handoff?

Handoffs happen for a few different reasons: a task exceeds one agent's specialization and needs a different agent better suited to the next step, a session needs to continue across a longer time span than one agent invocation covers, or the agent hits a limit — a low-confidence decision, a guardrail, an approval gate — and needs a human to take over. In every case, the quality of the handoff depends on how much relevant context transfers: a poor handoff drops history and forces the recipient to reconstruct what already happened, while a good one carries forward the conversation, the decisions already made, and why. This is harder than it sounds because context has to be summarized, not dumped wholesale — an agent or human receiving a handoff after a hundred prior steps needs the load-bearing facts, not the full transcript. Systems that do this well maintain a persistent session or memory object that any agent (or the escalated human) can pick up from, rather than reconstructing state from scratch each time.