What Actually Makes an AI "Agentic"? (Cutting Through the Hype)
Adwait Changan opens a 10-week Mycroft architecture series by defining agentic AI precisely: not a smarter model, but the same model wrapped in a loop with state, feedback, and a stopping rule.
"Agentic" has become one of those words that gets attached to almost anything with an API key and a spinner. A support bot is agentic. A search box is agentic. When a word covers everything, it stops explaining anything. This piece, opening a 10-week architecture series that eventually leads to the model context protocol, starts by stripping the word back down to something precise.
What a language model actually does, stripped down
Before building anything, it helps to look at what a language model does on its own, with nothing added. Text goes in, it predicts what comes next, text comes out, then it stops. It has no hands, no clock, and no memory of the last answer unless that answer gets pasted back in. That isn't a criticism of the model; it's simply the specification. Ask a plain chatbot to book a flight and you get a confidently written, numbered plan, but nothing was actually searched, no fare was checked, no seat was held. It handed you homework and called it help.
The one change that matters: wrapping it in a loop
The shift from chatbot to agent doesn't come from making the model bigger or smarter. It comes from wrapping it in a loop and giving that loop a way to touch the world. The loop works like this: the model gets a goal, it thinks and picks one next step, it acts by calling a tool that does something real, it observes what actually came back, and then it goes around again knowing something it didn't know before. This loop can be implemented in about 11 lines of code, and none of those lines are individually clever. The important part is the last line: a step budget. Without one, a confused agent can run forever, racking up cost the whole time.
Three additions: state, feedback, and a stopping rule
That loop hands a plain language model three things it never had on its own: state, meaning it remembers what it already tried; feedback, meaning the world answers back and that answer changes the next move; and a stopping rule. Two of those three sound like clear improvements. The third is the one that deserves the most caution, because it's also where the risk concentrates. A describer that gets something wrong wastes your afternoon. A doer that gets something wrong sends the email, deletes the row, or charges the card. Six failure modes arrive the instant that loop closes, and the worst of them is an agent that fails and reports success.
When an agent is actually the right tool
Not every task deserves an agent. The audit checklist comes down to three conditions: the task needs doing, not just describing; a wrong step is cheap to undo; and you can tell success from failure without auditing every step by hand. Miss that third condition and what you have isn't really an agent, it's a rumor about whether the task got done. The overall verdict is that an agent isn't a smarter model, it's the same model placed inside a loop that can call tools and read what comes back. The loop adds state, feedback, and a stopping rule, and that stopping rule is a design decision that belongs to the person building the system, not to the model itself. Agency is a trade, not an upgrade: the moment a wrong answer can become a wrong action, checking stops being optional.
Key takeaways
- Agentic AI is not a smarter or bigger model; it's a language model wrapped in a loop that can call tools and read the results.
- The loop adds three things a plain model lacks: state (memory of prior attempts), feedback (real-world results changing the next move), and a stopping rule.
- Six failure modes appear the moment the loop closes, and the most dangerous is an agent that fails while reporting success.
- Reach for an agent only when the task needs doing (not describing), a wrong step is cheap to undo, and success or failure can be verified without auditing every step by hand.
- Designing the stopping rule is a human responsibility, not something to leave to the model.
Try it yourself
This is the first episode of a 10-week Mycroft architecture series from Humanitarians AI Fellows, continuing next with a closer look inside a single turn of the agent loop. Take a task you do every week and ask an AI questions until you can describe it as an agent loop: the goal, the tools it would need, what it observes after each step, and the exact condition where it should stop and ask you. Then ask it to tell you honestly whether it should be an agent at all.
Chapters
- 0:00Cutting through the "agentic" marketing noise
- 0:40Why a standard chatbot hands you homework instead of helping
- 1:15Inside the 11-line loop: Bringing the model to life
- 1:50The three additions of a loop: State, Feedback, and Stopping Rules
- 2:30Six failure modes that arrive the moment you close the loop
- 3:10The audit checklist: When should you actually use an agent?
Full transcript(auto-generated, with timestamps)
Cutting through the "agentic" marketing noise
[0:00]This is onyx in for humanitarians AI. Fellow Adidweight Changen starts a 10-week series today and it ends at the model context protocol. It starts here with a word that has been stretched until it means almost nothing. Agentic. Every product is agentic. Now, so before we build a single thing, we are going to define it and the definition is smaller and stranger than the marketing suggests. Right now, agentic gets attached to anything with an API key and a spinner. A support bot is agentic. A search box is agentic. When a word covers everything, it explains nothing. So, here is the test that actually separates them. Strip a language model down to what it actually does. Text goes in, it predicts what comes next. Text comes out, then it stops. It has no
Why a standard chatbot hands you homework instead of helping
[0:41]Hands, no clock, and no memory of the last answer unless you paste it back in. That is not a criticism. That is the specification. So, ask a chatbot to book a flight and you get this. A numbered plan confidently written. Read it again. Nothing was searched. No fair was checked. No seat was held. It handed you homework and called it help. Now change exactly one thing. Do not make the model bigger. Do not make it smarter. Wrap it in a loop and give the loop away to touch the world. That is it. That is the whole trick. Here is the loop. The model gets a goal. It thinks, picks one next step. It acts, calls a tool that does something real. It observes, reads what
Inside the 11-line loop: Bringing the model to life
[1:16]Actually came back. Then it goes around again knowing something it did not know before. This is that loop as code. 11 lines and nothing clever in any of them. Pick a step if we are done. answer. Otherwise, call a tool and keep the result. Now, look at the last line. A step budget. Without it, a confused agent runs forever and bills you for the privilege. That loop hands the model three things it never had. State. It remembers what it already tried. Feedback. The world answers back and the answer changes the next move. And a stopping rule. Two of those sound like features. The third is the one that should keep you up at night. And here is
The three additions of a loop: State, Feedback, and Stopping Rules
[1:50]The bill. A describer that is wrong waste your afternoon. A doer that is wrong sends the email, deletes the row charges the card. Six failure modes arrive the instant you close that loop and the worst one is the last. An agent that fails and reports success. So when should you actually reach for one? Three conditions. The task needs doing, not describing. A wrong step is cheap to undo. And you can tell success from failure without auditing every step by hand. Miss that third one and you do not have an agent, you have a rumor. So the verdict, an agent is not a smarter model. It is the same model placed inside a loop that can call tools and read what came back. The loop adds state feedback and a stopping rule. And that stopping rule is yours to design, not
Six failure modes that arrive the moment you close the loop
[2:30]The models. Agency is a trade, not an upgrade. The moment a wrong answer can become a wrong action. Checking stops being optional. Reach for an agent when the task needs doing. When undo is cheap. And when you can verify the result without reading every step. Your turn. Here is the prompt. Take a task I do every week. Ask me questions until you can describe it as an agent loop. the goal, the tools it would need, what it observes after each step, and the exact condition where it should stop and ask me. Then tell me honestly whether it should be an agent at all. Run that on something you actually do. The last sentence is the interesting one. You are asking it to argue against itself. And watch how hard it is to name the stopping condition. That was what makes an AI agentic. Episode one of 10. Next
The audit checklist: When should you actually use an agent?
[3:12]Friday, we go inside a single turn of that loop and find the part that is actually loadbearing. Onyx in for humanitarians AI.
More from Humanitarians AI Fellows
3:25Inside an AI Agent Loop: The Only Step That Touches Reality
2:06Why Correlation Doesn't Equal a Tradable Signal (Spurious Correlation)
3:52How to Prove What Your Financial AI Actually Did (Auditing the Chain)
2:13Why Backtesting on Revised Financial Data is "Look-Ahead Bias"
5:29Quantum Sphere: The Extinction Paradox That Defies Everyday Physics
7:55