Cross-model handoff

Transferring a task, artifact, evidence, and state from one model or AI system to another during a workflow.

A handoff may use a different model for cost, latency, access, modality, independence, or a measured task advantage. It also introduces context loss, format translation, duplicated reasoning, and inconsistent instructions. A capable primary model can often complete several stages more coherently, so additional models should earn their place through evidence rather than a fixed draft-research-polish hierarchy.

Builder example

The receiving system needs the goal, current artifact, source evidence, constraints, completed work, unresolved questions, and acceptance checks. Copying only the latest prose can discard decisions and provenance. Use a shared structured record or generated handoff package when the transition repeats.

Common confusion: A cross-model handoff moves one artifact through several assistants in sequence so each contributes a stage. Running the same prompt in two assistants to compare their answers is a side-by-side test, where you pick one winner rather than combining their work into a single output.