- fractionl.ai
- Posts
- AI Agent Mythbuster
We’re all in agentic heaven at the moment, and kind of like every obsession before it (data, automation, internet, AI) there are going to be a lot of pretenders, bad actors and misinterpretations along the way.
So here is a really quick glossary and primer on what actually is what with all the concepts surrounding AI agents. Being a linguist by nature and qualification, I believe that dictionary definitions are usually the first step to alignment, understanding and good-natured adoption of new things. Let’s go.
Agent: It must be goal-driven, and it must be able to work while you sleep. It’s a software system that autonomously takes actions toward a defined goal using memory, reasoning, and environment feedback; not just responding to a one-off prompt.
Not just “a chatbot that books stuff.” An agent right now is more like a digital intern with initiative, if your intern had amnesia and occasionally hallucinated. Agents are going to improve hugely in the coming months, and be able to work with other agents like employees work with other employees.
Workflow Automation: A sequence of predefined tasks strung together by tools like Zapier, Make, or Airplane.dev. Trigger-based, deterministic, and rarely self-correcting.
It does exactly what you tell it to, even if that means forwarding spam to your boss.
Prompt: A text-based instruction that guides an LLM (like ChatGPT) to generate a desired output. The better the prompt, the better the response.
Crafting the perfect prompt is like trying to explain your feelings accurately to your spouse. A work in progress.
Multi Context Protocol (MCP): a prompting pattern where a model is called multiple times in a loop or sequence to refine its own output — often through planning, reflection, or step-by-step reasoning.
MCP is about thinking: the model is asked to think in phases, like a chess player planning several moves ahead.
Chain: A linked series of prompts or model calls where the output of one becomes the input to the next.
Basically a fancy workflow, but now with the risk of hallucinations along the way that need to be understood and/or managed.
Tool Use: The model's ability to call external APIs, databases, or functions to complete a task e.g., running code, browsing, or hitting a company API. It’s about action: the model “reaches out” beyond text generation to do something in the real world.
It’s what makes ChatGPT more like R2-D2 and less like Clippy in Microsoft Word.
Memory: Persistent storage of interactions or facts that an agent can recall later, enabling continuity across tasks or conversations.
Without it, every agent is a shit first date on repeat.
Planning: The agent's capacity to break a complex goal into sub-tasks and sequence them logically before acting.
Think of a really organized project manager you’ve worked with that has scattered attention or memory. Great potential, needs supervision or frameworks to fulfil their potential.
Reflection: The agent’s ability to evaluate past actions and improve its strategy or output over time.
Basically, “oops, I messed that up” but in code. You all know someone who never admits they’re wrong - it's irritating right?
Autonomy: The ability of an agent to operate independently, selecting actions and tools without direct user input at each step.
If you need to press “Enter” after every step, it’s not autonomous.
Orchestration: The infrastructure layer that coordinates multiple agents, tools, and environments to complete complex workflows.
Like being a conductor of an orchestra filled with Jack Blacks. They would desperately need to be kept in time with one another for the output to sound coherent.
Here are some common myths I see flying around on LinkedIn right now…
Myth #1: “I’m building an agent with n8n.”
No, you’re building a workflow. It’s an Airtable zap that calls an API. A nice workflow but not an agent. Real agents reason, recall, and react - not just respond.
Myth #2: “I talk to ChatGPT. I use agents every day.”
You’re chatting with a very helpful autocomplete. An agent isn’t a conversation, it’s a goal-seeking system. If it doesn’t remember, plan, or take action without you micromanaging it, it’s not an agent. It’s an articulate Google search bar.
Myth #3: “This agent books meetings for me!”
If it’s glued together with Calendly, Zapier, and some JSON, that’s an automated workflow, not an agent. You see this one all the time on LinkedIn when you’re told to reply AGENT in the comments to receive someone’s PDF. Not saying it’s not helpful. But it’s not agentic. Real agents negotiate tradeoffs, handle situations that are outside of the norm, and reroute or take a new approach when plans break.
Final thought
We need clarity and dictionary definitions if we are to not go down the same road we’ve gone down with our usage of people’s data over the last 20 years. With data, we're trying to unpick a world where everyone pays a lot to implement hyped up technologies, businesses say they’re doing great things and the system gets gamed (purposely by opportunists/scammers or by accident). That created a world in user data usage where promises weren't fulfilled, errors needed correcting and consumers hated your output. We can avoid the same with AI Agents if we all upskill on accurate definitions and myth busting. An agent isn’t just a prompt with plugins. It’s a system that works towards goals and does work while you sleep.
My aim is that we all mythbust continually so we can fix the problems at hand and not continue to wash dishes while our arm is on fire.
Basically, don’t be this guy.
Daily News for Curious Minds
Be the smartest person in the room by reading 1440! Dive into 1440, where 4 million Americans find their daily, fact-based news fix. We navigate through 100+ sources to deliver a comprehensive roundup from every corner of the internet – politics, global events, business, and culture, all in a quick, 5-minute newsletter. It's completely free and devoid of bias or political influence, ensuring you get the facts straight. Subscribe to 1440 today.