One test decides if you need an AI agent: if you can write the steps down, you don't

Virtual_Hair_1987 · reddit · 2026-09-22

A Reddit long-form argues there's a single test for agent vs. workflow: if you can write the steps down before the run starts, it's a workflow; if the next action depends on the previous return and is unknowable in advance, you need an agent.

Definitional split: Citing Anthropic's Building Effective AI Agents — workflows orchestrate LLMs via predefined code paths, while agents dynamically direct their own processes. It's an architectural spectrum: start with the simplest solution, add complexity only when it demonstrably helps, and optimizing single LLM calls suffices for most apps.

The tradeoff: Harrison Chase's one-liner — workflows trade autonomy for predictability, agents trade predictability for autonomy — but reliably good outcomes require neither alone. Workflow complexity lives in code (branching, parallel edges); agent complexity is abstracted into natural language, so anything unstated in the prompt is undefined at runtime.

Documented failures: Drawing on Arize AI's analysis of millions of decision paths, the post highlights failure modes that argue for writing a strict spec first: hallucinated tool arguments (e.g., agent assumes userid when the schema requires customeruuid, returning zero rows silently) and inefficient trajectories like the "polling tax" (looping on status checks instead of waiting for a webhook).

Original post →

One test decides if you need an AI agent: if you can write the steps down, you don't

More from coding & agent