Handoff★

A handoff is the process of transferring task progress, files, and architectural decisions from a bloated chat session to a fresh one. It allows you to reset the token count and restore the model's Smart Zone focus without losing your project's history or decisions.

You run handoffs whenever you:

• Transition an agent from a planning phase (designing an interface) to an implementation phase (writing code).
• Fan out a massive project into parallel agent sessions (running multiple subagents at once).
• Reset a bloated conversation that has entered the "dumb zone."

Handoff Mechanisms: Artifact vs. Compaction

To transfer state between stateless sessions, developers use two primary mechanisms:

1. Handoff Artifact (Explicit/File-based):
   • Form: A markdown file saved to your workspace (e.g. handoff.md).
   • Pros: You can read, correct, and edit the file before loading it. Reusable across multiple parallel sessions.
2. Compaction (In-context/Automatic):
   • Form: A summary generated by the model inside the chat history, used to seed the next request.
   • Pros: Quick and cheap, but lossy and harder for the developer to inspect.

The benchmark of a successful handoff is: Could a developer (or model) with absolutely zero context read the handoff material and continue the work correctly? If the successor session must ask you "What files did we edit?" or "What was the db structure we chose?", the handoff failed.

Field Applications

1. AI Engineers (Automating Handoffs in Multi-Agent Loops)

AI engineers structure agent pipelines to write a status file upon sub-task completion, triggering a cleanup script that starts the next agent container using that status file:

• Pipeline Flow: Agent A (Planner) writes todo.json ──► Container exits ──► Harness launches Agent B (Builder) mounting todo.json.