Glossary
what is an agent handoff?
TL;DR
A handoff is passing a task from one AI agent to another, or to a fresh session of the same agent, with enough context for the new one to carry on. You hand off when an agent's context window is full, when a different agent suits the next step better, or when an orchestrator gives a subtask to a worker. The hard part is deciding what to pass along: too little and the new agent repeats work, too much and it starts already cluttered.
Agents don't share memory. When one stops and another starts, everything the first one learned is lost unless it's written down and passed on. A handoff is that written note.
Why hand off
A full context, a better agent for the next step, or a task worth delegating.
Long sessions degrade as an agent's context window fills. Starting a fresh session with a short summary often works better than pushing on. Other times you want a different agent for the next step, for example one model to plan and another to implement, or a separate agent to review code it didn't write.
How tools do it
Some pass the transcript, some pass a summary, some pass a written brief.
Approaches vary in what they carry over.
| Tool | How it hands off |
|---|---|
| Superset | "Continue with another agent" passes the last part of the conversation, or the saved transcript, to a new agent. |
| Synara | Switches the provider mid-task with a handoff recap, and supports forks and side chats on other providers. |
| Orca | "Continue in New Session" starts a fresh agent, the same or a different CLI, with a bounded handoff prompt. |
| HarnessRouter | Changes a session's agent in place, passing a handoff document. |
| Pane | A Session or any agent starts a new pane with a prompt or a prompt file through the runpane CLI, so the brief is a document you can read. |
What makes a good handoff
The goal, what's done, what's left, and where to look. Not the whole chat log.
A good handoff names the goal, what's already done and verified, what's left, the decisions that were made, and which files matter. It leaves out dead ends and chatter. Writing it as a file the next agent reads, rather than pasting a transcript, keeps it short and lets a person check it too.
Related
Sources
- Corrections
- If anything here is wrong or out of date, tell us and we'll fix it and update the date at the top. Open an issue on GitHub
- Sources
frequently asked questions
By Parsa Khazaeepoul, co-founder of Pane. .