# Workflows vs. Agents

**Track:** Building the AI-Native Product — AI for Entrepreneurship — complete (29)
**Framework / surface:** venture strategy
**Level:** Advanced
**Prerequisites:** Product Design Under Uncertainty

**In one line:** Predefined steps or self-directed loops — the simplest structure that works, architected to be promotable.

## Theory, aesthetics & inspiration

Anthropic's "Building Effective Agents" (2024) drew the load-bearing distinction: workflows orchestrate models through predefined steps, while agents direct their own process — choosing tools and looping until done, the pattern research formalized as reasoning-plus-acting (ReAct, 2022). The engineering guidance is conservative and correct: use the simplest structure that works, because a workflow's predictability is a feature you surrender only when the task genuinely requires open-ended judgment about what to do next. The strategic guidance points the other way: as capability compounds, the frontier of tasks trustable to agents expands monthly, so architect workflows you can promote to agents without rewriting the product — the same decomposition, evals, and tool interfaces serve both, and the difference becomes a dial rather than a rewrite. The test for any given task is the question this lecture leaves you with: does it require judgment about the next step, or execution of known steps? Bill both accordingly, and re-ask the question every release.

**Founder question:** Does this task require judgment about the next step, or execution of known steps?