Run LLM calls in parallel — with typed output
Independent LLM work shouldn’t run one call after another. open-multi-agent runs the independent agents at once and merges their typed results.
The problem.
The naive shape crams summary, extraction, and classification into one giant prompt — or fires the calls sequentially, so latency is the sum of every step. Both waste the fact that the sub-tasks don’t depend on each other.
How open-multi-agent does it.
You hand runTeam() a goal; the coordinator decomposes it into a task DAG and runs the independent nodes in parallel, so wall-clock time is the slowest branch, not the sum. Each specialist can return Zod-typed, schema-validated output, and an aggregator merges them into one typed result — no manual Promise.all plumbing, no untyped string-stitching.
When this fits.
This fits when your task splits into independent sub-tasks that each want their own prompt (and often their own model), and you want typed output rather than one prompt’s best guess. If the steps are strictly sequential and dependent, a single agent is simpler.
Taking this to production?
open-multi-agent is MIT-licensed and free to run yourself. When you need it delivered, integrated, or supported on a deadline, 元定义科技 (YuanASI) offers commercial delivery and support.