Gating Cognition: How Modern AI Architectures Minimize Runtime Reasoning
summary
Compares observed ReAct agent flows that incur four or more LLM calls and tool descriptions in every prompt with a ReWOO planner that outputs a JSON plan handed to a deterministic executor so only two LLM calls happen. Demonstrates a chatbot called Osimo that orders "killer laser robots" and measures ~40,000 tokens and 19–20 second waits for the React demo versus ~21,000 tokens and much less blocking for ReWOO, arguing concurrency, lower cost, and higher reliability at the cost of flexibility for open ended reasoning.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:12Agentic architectures vs React RWO
- 3:42React architecture with LLM calls
- 8:58ReWOO reasoning without observation
- 14:41React vs ReWOO differences
- 15:02React demo with Osimo
- 18:41ReWOO demo with Osimo
- 23:24Performance and token efficiency
- 24:27Optimization strategies
- 25:08React vs ReWOO choice
related talks
19mRedefining Intelligence: How Reasoning Is Re-Shaping AI in 2025
01h 39mMastering LLMs: From Prompt Engineering to Agentic AI
52mThe Agentic Shift- Are We Reasoning, Or Just Retrieving Smarter?
29mLLMs that think: Demystifying Reasoning Models
25mFrontier AI Models: A Review of Past Nine Months
46mDemystifying AI Agents
24mRecursive Coding Agents
34mAI Agents: Transforming Autonomous Decision-Making