Not every multi-step AI task needs an agent — workflows still win for predictable paths

Old-Farmer-1029 · reddit · 2026-09-10

A Reddit discussion on whether every multi-step AI workflow actually needs to be an agent. The author argues many systems pile on multi-agent setups, memory, and planners when the task path is actually fixed — that's a workflow, not an agent. Agents shine when the path isn't known in advance: investigation, choosing tools dynamically, recovering from surprises. For predictable processes, deterministic workflows with AI only where it adds value are preferable. The post also stresses the demo-vs-production gap: retries, state, permissions, approvals, logging, and auditability suddenly matter. It closes by asking readers where they draw the line, and whether they've ever replaced an agent with a simpler workflow.

Original post →

More from coding & agent

coding & agent channel →