Free PDF · ~12 pages
AI Agent Loop Architecture
When to use ReAct vs. Plan-Execute vs. Reflexion. The decision tree we use at Qyndex, in printable form. Picking the wrong loop wastes 5x the tokens; this is the cheat sheet that stops it.
What’s inside
- The three loop shapes side-by-side. ReAct (think-act-observe), Plan-Execute (planner + cheap workers), Reflexion (self-critique with iter cap). One paragraph each, with the loop shape in pseudocode.
- The decision tree. Latency budget → task structure → quality bar. Five branches, three terminals, no ambiguity.
- Real cost numbers. Tokens, USD, p95 latency for each shape on a Qyndex content campaign. Plan-Execute saves 54% vs. ReAct on bulk; Reflexion adds 46% but catches the 8% of brand-voice drifts.
- Which we use where. How the Qyndex supervisor graph (Researcher / Strategist / Writer / QC) wires all three. Reusable as a template for your own multi-agent stack.
Who it’s for
- Engineers picking an agent framework and trying to skip the trial-and-error
- Tech leads sizing the cost / latency envelope before the team commits to a shape
- Founders shipping LLM features who want a defensible "why this loop" answer