Hard problems don't fit in a single prompt. Knitch gives you the orchestration patterns, decompose, fan out, evaluate, route, as node types you drop on the canvas. No queues to manage, no glue code to babysit.
Each of these is a node type you drop in and wire up, not infrastructure you build and maintain.
An orchestrator breaks a task into subtasks, workers handle them in parallel, and an aggregator merges the results. The flagship pattern, in three nodes.
Run several AI nodes at once and merge what comes back. Good for trying a few approaches, or processing independent inputs side by side.
The orchestrator decides how many workers to spawn at runtime, one per task in its output. Ten items in, ten workers, nothing hardcoded.
A generator drafts, an evaluator scores it against your criteria, and the loop refines until it passes or hits a limit. Verdicts are pass, revise, or escalate.
A router classifies the input and sends it down the right branch, so each case gets the model and the prompt that actually fit it.
Pause a run for a person to approve, edit, or escalate, then let it continue from where it stopped. The agents don't have to decide everything alone.
It runs in parallel, for real. Every worker is its own durable task. Knitch fans them out, runs them at once with retries, waits for all of them, and guards against runaway nesting. You watch each agent execute live, and you see what every one of them cost.