Four Levels: Chatbots to Agentic AI Systems

Simon Scrapes2026-05-09go watch the original →

the gist

Video demystifies agentic AI for non-devs via content repurposing example: level 1 chatbots give static advice; level 2 workflows automate fixed steps (n8n/Zapier); level 3 agents decide paths in harnesses (Claude Code/Cursor via ReAct); level 4 systems coordinate skills, memory, MCPs with human-in-loop.

Progression from Static to Autonomous AI

Simon Scrapes outlines four levels of AI automation for non-developers, using YouTube video to social posts/newsletter repurposing as a consistent example. Level 1 chatbots (ChatGPT, Claude, Gemini) provide passive advice with static context like voice guidelines or projects/gems; they require manual prompting and lack business knowledge or action. Level 2 AI workflows (n8n, Zapier, Make.com) automate fixed pipelines: pull transcript, prompt Claude with hardcoded guidelines, output to scheduler; they repeat steps without adapting to trends like carousel performance or platform fit.

Agentic Workflows Add Decision-Making

Level 3 introduces agentic workflows in harnesses (Claude Code, Cursor, Codex), where a single agent handles one goal like "Turn this video into LinkedIn/Twitter/newsletter content." The agent pulls transcripts, assesses topics against viral trends, drafts platform-specific formats (e.g., LinkedIn carousel for visuals, X thread for contrarian angles), applies style guides, and saves for review. This uses a ReAct loop: the model reasons, acts (reads files, runs commands), observes, and iterates. Harnesses enable file access, tool calls, and self-checking, unlike browser chatbots.

Agentic Systems Coordinate Operations

Level 4 scales to multi-agent systems for full operations: one trigger runs clip extraction, carousel building (platform dimensions/brand aesthetics), newsletter drafting, ad copy from past winners, and scheduling queuing. Building blocks include skills (task folders with instructions/examples loaded on-demand), memory (markdown file or DB tracking post performance/open rates across sessions), MCPs (Model Context Protocol for tool plugs like schedulers/analytics/CRMs), and deliberate human-in-loop (e.g., output review before publish). Tools like his Agentic OS community build or open-source OpenClaw/Hermes layer files/folders atop base harnesses; everything reduces to editable files, akin to Notion workspaces, no code needed.