Recipes

Recipes are reusable, parameterized workflows. Each compiles to a DAG of motes and inherits the runtime's guarantees — exactly-once delivery, transparent recovery, and policy gating — for free.

map_reduce

N mappers fan out, then a single pure reduce folds their results.

input → [mapper × N] → reduce → output

Use it for embarrassingly-parallel work with a deterministic aggregation step.

fan_out_gather

N parallel workers, then one pure gather. Like map_reduce but the gather collects heterogeneous worker outputs rather than folding a uniform list.

input → [worker × N] → gather → output

retry_until_critic

N attempts, each gated by a critic, then a selector picks the winner. The critic decides whether an attempt passes; failing attempts are retried up to the bound.

attempt → critic ─(reject)→ attempt → … → selector → output

Use it when output quality is checkable and you want bounded retries instead of one-shot.

react_tool_loop

The classic reason → act(tool) → observe loop: the agent reasons, calls a tool through the CapabilityBroker, observes the result, and repeats until done.

reason → act (tool) → observe → reason → …

image_batch_describe_reduce

Batch image description: describe a batch of images in parallel, then reduce the descriptions into a single result. A multi-modal variant of map_reduce.