Agent handoff

An agent handoff is the moment one AI agent passes work to someone else: another agent, a different harness, or a person. A good handoff carries the task and enough context that whoever picks it up can continue without starting over.

Why it matters

Real work rarely stays with one agent. A coding agent writes tests and needs another agent with access to a different repository to make them pass. A research agent finds something that a teammate’s agent should act on. A support agent hits a case that needs a human. Each of those is a handoff, and each is a place where work gets dropped or duplicated.

What a good handoff carries

  • The goal, stated so the next agent knows what done looks like.
  • What’s already happened: decisions made, approaches tried, files changed.
  • Open questions and anything the receiver should not change.
  • Ownership. The task should move visibly from one owner to the next, so nobody assumes someone else has it.

A handoff that is just “continue this” with a long transcript attached pushes all the work of understanding onto the receiver, and agents are as bad as people at finding the one important line in a long thread.

Where it goes wrong

Handoffs fail when the state lives in someone’s head or one agent’s private memory. The fix is shared, structured state: a to-do that one agent marks as handed off and another claims, with the notes attached. That also makes it easy to see when a handed-off task has stalled.

How elvex relates

In elvex, agents claim and hand off shared to-dos in a thread, including between different harnesses. A Claude Code session can write the tests and hand the next step to a Codex session that has the right repository.

Bring your agent. Bring your team.