Process multiple requests

First decide what “multiple” means in your application. A multiselect question returns several labels for one subject. A request_units question identifies separate requested actions, each with an ID, status, category, and optional subject. Two labels are not automatically two jobs.

For example, “Freeze card-123 and send the statement for account-9” can have both labels in multiselect; request_units can additionally represent the two distinct actions. A single-choice classification may correctly abstain with multiple_valid_options. The decision evidence example demonstrates those question types and their different outputs.

Plan before processing

A request-unit assessment is information, not authorization to execute each unit. Units can be active, withdrawn, conditional, or quoted, and the answer may contain explicit relations. The host's trusted policy should check the assessment status, supported categories, subjects, duplicates, relations, and any business prerequisites before creating a work plan.

The multi-request example uses assess → plan → process → finalize. Its trusted plan_requests handler creates ordered {id, flow, input} items only for allowed independent work; held and ignored units remain visible for final disposition. The process step uses a flow_collection with an explicit callable-flow allowlist and item cap.

type: flow_collection
items:
  pointer: /payload/items
flows:
  - lookup_status
  - prepare_guidance
max_items: 8

Every item is validated before child I/O. Callable flows run sequentially under the parent's admission slot, deadline, and budgets. A child review is recorded and later independent items continue. A technical child failure stops new work and preserves the ordered ledger in partial_result.items; later items are skipped. After a successful or review collection, read result.items and let trusted code determine the final business disposition.

Run the complete example offline with its scripted model and local tool:

uv sync --locked --extra mcp --extra openai
uv run --no-sync python -m examples.multi_request_processing.run

The default run does not call a model endpoint. See Configure a flow collection for the full item and ledger contract and score request-unit results before using the policy with real inputs.