Agentic AI Systems, Clearly Explained
This video explains agentic AI systems across four levels of complexity: chatbots, AI workflows, agentic workflows, and agentic AI systems. Using content repurposing as a consistent example, the speaker demystifies intimidating terms like harnesses, MCPs, and memory systems, arguing that these systems are fundamentally just files and folders accessible to non-developers.
Summary
The video walks through four progressive levels of AI capability using a single running example — repurposing a YouTube video into social posts, newsletters, and clips — so viewers can feel the difference at each stage.
Level 1 covers chatbots (ChatGPT, Claude, Gemini). The speaker notes that while useful, chatbots are passive and stateless: they don't know your brand, audience, or content history unless you manually paste that context in every time. Projects and Gems offer static context storage, but the AI still waits for prompts and doesn't execute anything autonomously.
Level 2 covers AI workflows (n8n, Zapier, Make.com). These tools automate repetitive pipelines — for example, automatically pulling a transcript, sending it to Claude with a hardcoded prompt, and dropping a draft LinkedIn post into a scheduling tool. The limitation is rigidity: the workflow follows the same steps in the same order every time and cannot make judgment calls, such as recognizing that a topic suits Twitter better than LinkedIn.
Level 3 introduces agentic workflows, which the speaker identifies as the most important conceptual leap. Here, the model — not the human — decides the execution path. Given a high-level goal like 'turn this week's video into content for LinkedIn, Twitter, and my newsletter,' a tool like Claude Code reasons about what to do, acts, observes results, and iterates. This loop is technically called ReAct (Reason and Act). The infrastructure enabling this is called a harness — Claude Code, Codex, and Cursor are all described as harnesses that wrap the model and give it the ability to read files, run commands, and call tools.
Level 4 describes full agentic AI systems: coordinated multi-agent operations where different 'skills' (markdown instruction files) handle specific tasks like clip extraction, carousel creation, newsletter drafting, and ad copy generation. The system loads only the relevant context at the right time to avoid bloating the context window. Memory — which can be as simple as a markdown file — allows the system to carry knowledge across sessions, such as which post formats performed best last month. External tools connect via MCP (Model Context Protocol). Critically, the speaker emphasizes that human-in-the-loop review is a deliberate design principle, not a limitation — nothing gets published without human sign-off. The speaker concludes that despite the intimidating terminology, these systems are essentially organized folders of files, making them accessible to business owners and knowledge workers, not just developers.
Key Insights
- The speaker argues that the critical distinction between AI workflows (Level 2) and agentic workflows (Level 3) is who decides the execution path — in Level 2 the human defines the steps, while in Level 3 the model determines them based on a high-level goal.
- The speaker defines a 'harness' as the infrastructure surrounding a model — such as Claude Code, Codex, or Cursor — that turns thinking into doing by giving the model access to files, the ability to run commands, and connections to external tools, distinguishing it from a browser-based chatbot.
- The speaker claims that 'skills' in a Level 4 agentic system are simply markdown files containing instructions for specific tasks, and that the agent loads only the relevant skill and reference examples when needed to avoid bloating the context window and wasting tokens.
- The speaker asserts that memory in agentic AI systems can be as simple as a markdown file the system reads and updates between sessions, allowing it to remember things like which post formats performed best last month rather than starting from zero every time.
- The speaker states they have not yet encountered a system powerful enough to bypass the human-in-the-loop review stage, and that the best real-world agentic systems are deliberately designed with human review built in at either the input or output stage before anything goes live.
Topics
Full transcript available for MurmurCast members
Sign Up to Access