-
Notifications
You must be signed in to change notification settings - Fork 0
DynamicWorkflowDesignPatterns
title: Dynamic Workflow Design Patterns radar_quadrant: Techniques radar_ring: Assess radar_position: inner
Teams building systems that coordinate many AI agents at once keep reinventing the same handful of coordination shapes: splitting work across many workers and merging the results, having one agent check another's output, repeating a step until it stops producing improvements. A field note by zircote catalogs twenty of these recurring shapes under a shared vocabulary, arguing that "the community has already converged on a handful of recurring shapes... and those shapes have names in scattered blog posts but no systematic treatment."
The catalog groups patterns into four families: structural (how work is spatially arranged, such as fan-out and pipelines), behavioral (how work unfolds over time, such as loops and staged execution), verification (how confidence is manufactured, such as running a skeptic panel or cross-checking results), and contract (how data boundaries are enforced between agents, such as schemas and parameter validation).
Two organizing principles run through the catalog. First, the structure of how work is arranged matters more than how many agents are involved. Second, each pattern is presented as a named trade-off rather than a ready-made recipe — the note is explicit that adopting a pattern means accepting its documented costs, not getting something for free.
A representative example is the first pattern, fan-out/reduce/synthesize: given nine independent sources to check, one researcher is assigned per source running in parallel, each returning schema-validated results; a small script flattens and deduplicates the output; one curator then ranks the combined results. The pattern separates deterministic control flow (handled by ordinary code) from stochastic work (handled by individual agents), keeping intermediate results out of the main conversation.
This is placed in Techniques because the field note documents a reusable design vocabulary for orchestrating AI agent workflows, not a specific product. It sits in Assess because the value is in the external maturity and systematic treatment of a previously ad-hoc practice — there is no confirmed first-person production use of this pattern catalog, which is required to justify Trial or Adopt. Position is inner, reflecting that the catalog is well-organized and immediately applicable as a reference, even without production track record.