Where does the agent end and plain code begin? One dev's rule for splitting workflows

krunal_builds · reddit · 2026-11-20

A developer sparks a discussion on drawing the line between agent judgment and deterministic code: in a typical workflow, only 4 steps involve real judgment (e.g. parsing messy customer emails) while 8 steps are just lookups, validation and writes.

His working rule: anything with a checkable right answer stays in plain code, and the model only handles messy inputs. The system becomes boring and cheaper to run, but the agent portion looks small — which some clients read as under-delivering.

He asks whether others let the model orchestrate end-to-end with guardrails, or keep orchestration deterministic and call the model at specific points.

Original post →

More from coding & agent

coding & agent channel →