Cognitive Orchestration
Enroll

Free preview · Lesson 7

Agentic AI & AI Agents

At the top of the stack sits agentic AI: systems where the model does not just answer — it directs its own process, deciding which tools to call and in what order to reach a goal. This is the home ground of the runtimes this course is built on: Hermes Agent runs as a long-lived agent on your own server — persistent memory, autonomous skill creation, subagent delegation, and a Kanban board where named agents collaborate — while OpenClaw is a personal agent you message on WhatsApp, Telegram, or Slack that controls your computer and works on its own schedule, even overnight. Anthropic's distinction is the cleanest mental model for reasoning about both: a workflow orchestrates LLMs through predefined code paths; an agent lets the model dynamically control how it accomplishes the task. With a workflow, you own the plumbing; with an agent, the model owns it.

What you'll take away
  • Describe the agent loop — reason, act, observe, reflect — and the ReAct pattern.
  • Weigh the agent trade-off: flexibility against latency, cost, and predictability.

At the top of the stack sits agentic AI: systems where the model does not just answer — it directs its own process, deciding which tools to call and in what order to reach a goal. This is the home ground of the runtimes this course is built on: Hermes Agent runs as a long-lived agent on your own server — persistent memory, autonomous skill creation, subagent delegation, and a Kanban board where named agents collaborate — while OpenClaw is a personal agent you message on WhatsApp, Telegram, or Slack that controls your computer and works on its own schedule, even overnight. Anthropic's distinction is the cleanest mental model for reasoning about both: a workflow orchestrates LLMs through predefined code paths; an agent lets the model dynamically control how it accomplishes the task. With a workflow, you own the plumbing; with an agent, the model owns it.

The single structural choice that defines everything else: who decides the next step.
The single structural choice that defines everything else: who decides the next step.

How an agent runs

The dominant pattern is the agent loop — reason, act, observe, repeat — formalised as ReAct (Yao et al., 2022), which interleaves reasoning traces with tool calls. Around this core, research added the capabilities that make agents practical: Toolformer (teaching models to call tools), Reflexion (learning from verbal self-feedback), and generative agents (long-lived memory and planning).

The reason–act–observe–reflect cycle. Goal, tools, and memory sit at the centre; stop conditions keep it bounded.
The reason–act–observe–reflect cycle. Goal, tools, and memory sit at the centre; stop conditions keep it bounded.

Example. Message OpenClaw "find the three cheapest flights and book the one that lands before noon." It plans (search first), acts (calls a flight tool), observes (reads results), reflects (none land before noon — widen the search), and loops — without you scripting each branch. Hermes Agent does the same on a server, and can hand parts of the job to subagents on its Kanban board while you sleep.

The discipline this course teaches

That power is exactly why discipline matters. The ClawHavoc supply-chain campaign — over a thousand malicious skills planted in OpenClaw's marketplace to steal credentials — is the cautionary tale the security lessons return to: an agent that can act is an agent that can be made to act against you. Anthropic's central advice is equally sobering: find the simplest thing that works, and only add agency when the task demands it — agents trade latency, cost, and predictability for flexibility. That single trade-off is why every layer beneath this one still matters, and why the fourteen lessons ahead treat orchestration as a discipline of restraint as much as capability.

References & further reading

Workflow Automation (n8n, Make, Zapier)First Principles and Problem Anatomy
Resources