Concepts

Cori is built on one insight: the LLM's value is figuring out the workflow on the first conversation — not re-deriving it on every run.

When you solve a problem with an agent, the agent does real thinking: it breaks the task into steps, picks tools, handles edge cases. That thinking is expensive and slow. But once the plan is known, you don't need the LLM to re-execute it. You need a reliable runtime.

Cori makes this split explicit: the agent does design-time reasoning; the workflow runs deterministically at runtime on Temporal. An LLM is only in the loop at runtime if a step genuinely needs to process new data — to translate, classify, summarize, or extract something from content that didn't exist at design time.