What Actually Makes an AI "Agentic"? (The 4 Tiers)

Agentic AI runs on a spectrum, not a yes-or-no label, and running one task, booking a flight, through four tiers of capability shows exactly where most products marketed as agentic actually land.

4:51 video5 min readWatch on YouTube

"Agentic" shows up on every startup's landing page, every keynote slide, every LinkedIn post promising the future has arrived. The problem is that half the products calling themselves agentic are doing something a form on a website could basically do, while the other half are running loops that can rack up real costs or real consequences while nobody's watching. Rather than argue over a dictionary definition, one Humanitarians AI Fellow runs a single task, booking a flight, through four tiers of AI capability, giving a concrete way to tell where any agentic claim actually lives.

Tier zero: a single question, a single answer

At the floor of the spectrum, tier zero is a single question and a single answer. You type "book me a flight to Chicago next Friday," and the AI replies with words, maybe something like "try Google Flights, compare a few times, book whichever fits your schedule." That's helpful advice and zero action. Underneath, the model is just predicting the next likely word over and over until it's built a full sentence. There's no flight search, no browser, no bank card involved. It's the equivalent of texting a well-traveled friend for advice: great advice, but they didn't lift a finger to book anything.

Tier one: a single tool call

Tier one is where things start becoming real. The model now has access to a tool, say a flight search function. Ask the same question, and this time it doesn't just talk about flights, it sends a request to a real flight search system, gets back live prices and times, and shows actual options. This is what people usually mean by function calling or tool use: the model recognized it needed outside information, formatted a request the tool could understand, waited for the result, and used it. But it stops there. It hands you a list, and you still pick a flight, enter your card, and hit purchase. One tool call, one job done, then control comes straight back to you.

Tier two: planning across multiple steps

Tier two is where an AI stops acting like a search bar and starts acting more like an actual assistant. Instead of one tool call, it chains several together, building the plan itself as it goes. Ask something more complex, "book me a flight to Chicago next Friday, but only if it doesn't clash with anything on my calendar," and the model has to actually check the calendar, search flights, line the two up against each other, notice a conflicting 9 a.m. meeting, and come back with something like "there's a 7 a.m. flight that works, or an 11 a.m. one that would make you miss your meeting, which do you want?" That's planning: a fuzzy goal broken into ordered steps, multiple tools used in sequence, and the plan adjusted based on what it actually found. This is the tier most products marketed as AI agents genuinely live in. It looks autonomous, but it's bounded, a handful of steps and then it checks back in before spending your money.

Tier three: the autonomous memory loop

Tier three is the one everyone actually means when they say "agentic," even though most products on the market can't really do it yet. This is an AI that doesn't just run a plan once and stop. It runs in an ongoing loop, watching, deciding, acting, and remembering, working toward a goal over hours, days, or weeks without anyone standing over its shoulder at every step. Picture saying, once, "get me the best fare to Chicago sometime this month, book it the moment the price looks good." Tier three means it checks prices every day on its own, remembers that you like aisle seats and morning departures because you mentioned that in a completely different conversation three weeks earlier, notices the airline canceled your original flight, rebooks you automatically, and only messages you afterward to say it's handled. Nothing restarted from zero. It's carrying context and decisions forward across time and making its own calls about when to act. This is genuinely powerful, and genuinely risky.

The three-question audit checklist

The tear-down produces a simple checklist for evaluating any "agentic" claim on a landing page. First, does it use tools, or does it just talk? If it's just talk, it's tier zero, no matter how it's branded. Second, does it chain multiple steps into a plan, or does it fire off one tool call and stop? Chaining multiple steps is tier two behavior. Third, does it keep running without you, and does it remember what happened in past sessions? That's the only real test for tier three. Most of what's sold as agentic today comfortably sits in tier one or tier two. True tier three, autonomous, memory-carrying loops, is rare, harder to control, and still being figured out even by the best labs building this technology.

Key takeaways

  • Agentic AI is a spectrum, not a yes-or-no label, ranging from tier zero, talk with no action, to tier three, autonomous loops with memory.
  • Tier one means a single tool call with control handed straight back to the user; tier two means multiple chained steps with a check-in before spending money.
  • Tier three requires ongoing autonomous action, memory carried across separate sessions, and decisions made without a human standing by at each step.
  • Most products marketed as "agentic AI" today actually operate at tier one or tier two.
  • A genuine tier three claim needs to show it names the specific tools it calls, runs a plan across more than one step, and proves memory carries across separate sessions.

Try it yourself

Audit an AI product you actually use: which tier is it really, zero, one, two, or three, and what evidence proves it? Check the answer against whether it names specific tools, shows a multi-step plan, and proves memory persists across sessions, work in the spirit of the Humanitarians AI Fellows program's hands-on AI evaluation.

Chapters

  1. 0:00Cutting through the "agentic" marketing hype
  2. 0:45Tier 0 & Tier 1: From helpful advice to single tool calls
  3. 1:30Tier 2: Multi-step planning and adapting on the fly
  4. 2:15Tier 3: The holy grail of autonomous memory loops
  5. 3:00The 3-question audit checklist for any landing page
Full transcript(auto-generated, with timestamps)

Cutting through the "agentic" marketing hype

[0:00]Hi, I'm Divage Pow. This video is about what actually makes an AI agentic and where that label stops being true. You've seen the word on every startup's landing page, every keynote slide, every LinkedIn post promising the future just showed up. But here's the problem. Half the products calling themselves agentic are doing something a form on a website could basically do. And the other half are running loops that can rack up real costs or real consequences while you're not even watching. So today, we're doing a tear down. Not a dictionary definition, a spectrum. We're taking one task, book me a flight, and running it through four tiers of AI capability. By the end, you'll be able to look at any agentic claim and know exactly where it actually lives. Let's start at the floor. Tier zero is a single question, a single answer. You type, "Book me a flight to Chicago next Friday." The AI

Tier 0 & Tier 1: From helpful advice to single tool calls

[0:46]Replies with words, maybe something like, "Try Google flights, compare a few times, book whichever fits your schedule. Helpful advice, zero action. Here's what's actually happening under the hood. The model is predicting the next likely word over and over until it's built a full sentence. That's the entire mechanism. There is no flight search, no browser, no bank card. If this were a person, it's like texting your welltraveled friend for advice. Great advice, but they didn't lift a finger to book anything. Tier one is where things get real. Now, the model has been given access to a tool, say a flight search function. You ask the exact same question, and this time it doesn't just talk about flights. It sends a request to a real flight search system, gets back live prices and times, and shows you actual options. This is what people mean by function calling or tool use. The model recognized it needed

Tier 2: Multi-step planning and adapting on the fly

[1:32]Outside information. Format a request. The tool could understand, waited for the result, and used it. But watch closely. It stops there. It hands you a list. You still pick a flight, enter your card, hit purchase. One tool call, one job done, then control comes straight back to you. Tier two is where the AI stops acting like a search bar and starts acting more like an actual assistant. Instead of one tool call, it chains several together, building the plan itself as it goes. Say it differently this time. Book me a flight to Chicago next Friday, but only if it doesn't clash with anything on my calendar. Now the model has to actually check your calendar, search flights, line the two up against each other, notice you've got a 9:00 a.m. meeting that same morning and come back with something like, "There's a 7 a.m. flight that works or an 11:00 a.m. one that

Tier 3: The holy grail of autonomous memory loops

[2:15]Would make you miss your meeting. Which do you want? That's planning. A fuzzy goal got broken into ordered steps. Multiple tools got used in sequence and the plan adjusted based on what it actually found along the way. This is the tier most products marketed as AI agents genuinely live in. It looks autonomous, but it's bounded. A handful of steps, then it checks back in with you before spending your money. Now, tier three, the one everybody actually means when they say agentic. Even though most products on the market can't really do it yet, this is an AI that doesn't just run a plan once and stop. It runs in an ongoing loop, watching, deciding, acting, remembering, and it keeps working toward a goal over hours, days, sometimes weeks without you standing over its shoulder at every step. Picture this instead. You say it once, "Get me the best fair to Chicago sometime this month. Book it the moment the price

The 3-question audit checklist for any landing page

[3:00]Looks good." Tier three means it checks prices every day on its own. It remembers that you like aisle seats and morning departures because you mentioned that in a completely different conversation 3 weeks ago. It notices the airline canceled your original flight, rebooks you automatically, and only messages you afterward to say, "Handled it. Nothing restarted from zero." It's carrying context and decisions forward across time and making its own calls about when to act. This is genuinely powerful and genuinely risky. So, here's your tear down checklist. Next time you see agentic AI slapped on a landing page, ask three questions. One, does it use tools or does it just talk? If it's just talk, it's tier zero, no matter how it's branded. Two, does it chain

Multiple steps into a plan or does it fire off one tool call and stop? Chaining is tier 2 behavior. Three, does it keep running without you? And does it remember what happened in past sessions? That's the only real tier three. Most of what's sold as agentic today comfortably sits in tier 1 or tier 2. True tier three autonomous memory carrying loops is rare, harder to control, and honestly still being figured out by the best labs building this stuff. So, here's the whole spectrum in one frame. Tier zero talks and touches nothing. Tier one makes a single tool call and hands control straight back. Tier two plans several steps, adapts, and still checks in before it spends anything. Tier three runs a loop with memory for

Days without you. Agentic was never a yes or no label. It's a spectrum of how much a system does on its own and how far it carries what it knows between actions. Now you know exactly where to look. Your turn. Here's the prompt. Audit an AI product I use. Which tier is it really? 0, 1, 2, or three? And what evidence proves it? Run that on a tool you actually pay for. Then check the answer against three things. Does it name the specific tools the product can call? Does it show a plan running more than one step? And does it prove memory carries across separate sessions? If a tier three claim comes back without that last one, the claim is marketing. Signing off. Divage power.

More from Humanitarians AI Fellows

Humanitarians AI Lyrical Literacy Project