Guide · Agent orchestration
Pass Context Between AI Agents Without Sending Everything
Build a small handoff packet with the goal, current artifacts, decisions, evidence, and unresolved questions.
A worker can receive thousands of words and still be missing the one decision that matters. Good context is the information needed for the next assignment, organized so the worker can find it. Start from what the worker must decide rather than from how much history you have available.
Include the current goal and revision
For a hypothetical article-production workflow, send the intended reader, the article's purpose, the approved outline, and the current draft. If the writer changed the draft after review began, the reviewer needs to know which version it is reading.
Use a stable artifact identifier or revision when possible. A path such as "latest-draft" can be convenient for people, but it is ambiguous when several workers read and write at different times. Record the exact version associated with an assignment and its returned result.
Carry decisions with their reasons
A short decision record can prevent repeated debates. For example, the team may have removed a pricing comparison because the source could not be verified. Tell the next worker that this is an editorial decision and point to the relevant note.
Distinguish a decision from an unresolved question. "Do not include an unsupported price" is a rule. "We have not confirmed the current price" is a fact about the work. Both may matter, but they should not become a vague instruction to avoid the whole subject forever.
Send evidence by reference
Keep large logs, documents, and research results available as artifacts. Put the key findings and exact references in the handoff. A worker can then inspect the evidence behind a claim without rereading every unrelated source.
Claude Code documents separate context for ordinary subagents, with forks as an exception that inherit conversation context. Check the behavior of the platform you use. Do not assume a worker can see a file merely because another agent already read it, or that every new worker starts with the same history.
Ask for a usable return packet
The return should explain the result, changed artifacts, checks performed, and remaining uncertainty. If the worker learned something that changes the next assignment, put that near the beginning. Keep incidental exploration out of the main summary.
- Goal and expected deliverable.
- Current artifact identifiers or source revision.
- Decisions and constraints that affect the assignment.
- Findings with links to supporting evidence.
- Open questions and a clear stopping condition.
Test the handoff with a fresh reader
Give the packet to someone who did not follow the original conversation, or use a fresh worker in a controlled trial. Watch which questions it has to ask before beginning. Those questions often reveal missing context more clearly than reviewing the packet yourself.
Update the handoff format from those observations. A useful packet can stay short, but it should make the next action possible without encouraging the worker to invent the background.