Handoffs are becoming a first-class pattern in Claude workflows. Here is how I have been thinking about them.
Long Claude sessions still break on context decay. Handoffs are the simple fix: compress what matters, start a fresh agent, keep going.
Matt Pocock's new handoff skill (repo) does this in one command. It compacts the conversation into a document, points at existing artifacts instead of restating them, and the next agent picks up from it. It also chains between threads: /grill-with-docs -> /handoff -> /prototype -> /handoff back.
I built handoffs into APM, a multi-agent framework for Claude Code, back in May 2025 (1 year ago....) when context windows were tiny enough that you had to constantly start fresh or you would have to deal w hallucinations all the time.
What I did differently: split the handoff into two artifacts.
- a persistent narrative file recording what was done and decided and why
- an ephemeral prompt telling the incoming agent how to rebuild context from the codebase and that persistent file
The incoming agent reconstructs from durable project state, not just the compressed chat conversation. Persisting the file also leaves a trail, so once more than one agent is involved and you deal with multi-agent systems, you can keep track of when one is working off a summary rather than firsthand context. Easier to manage context gaps better.
I opened an issue on Matt's repo with a few of these ideas: mattpocock/skills#235.
How do you handle handoffs? Manual summaries, a skill, subagents? And does the two-file split resonate, or is one document enough?
EDIT: In the frameworks docs I have a dedicated session explaining how handoff works there. It applies generally.. you can get ideas and apply them to Matt's skill. https://agentic-project-management.dev/docs/agent-orchestration#memory-and-project-state