AI Agent vs LLM vs Assistant: Wetin Be the Difference?

Wetin be the difference between AI agent, LLM and AI assistant?

AI agent, LLM and AI assistant na three layers for one stack. To tell dem apart, ask wetin each one need from server. LLM (large language model) na file wey contain weights. E need RAM and compute. Assistant na that model wey dem wrap inside chat surface, with account and saved history. Almost every time, e dey run for another person hardware. Agent na assistant wey get tools and loop too. E dey hold credentials. Na this one make e need machine wey dey on continuously.

Plenty writing about this question stop for definitions. The definitions matter because each layer get different cost. One dey cost RAM. The next one dey cost public URL and TLS (transport layer security). The last one dey cost credentials. Once agent don use credential, you need rotate that credential.

LLM na weights, and weights need RAM

LLM na file wey contain numbers. You download am, runtime load am into memory, and e answer one request at a time. The contract narrow: text go in, text go out. Model no get memory between calls, no clock, no network access, and no way to open file. Anything wey LLM appear to remember na the program wey call am paste am into its context.

The number wey decide your VPS plan na the size of that file, because the whole file dey memory while model dey run. For the 4-bit quantisation wey Ollama ship by default, count roughly 0.6 GB for every billion parameters. Then add one or two gigabytes for context window and runtime itself.

[
  {
    "label": "qwen3:4b",
    "download_gb": 2.5,
    "ram_gb_needed": 6
  },
  {
    "label": "qwen3:8b",
    "download_gb": 5.2,
    "ram_gb_needed": 8
  },
  {
    "label": "qwen3:14b",
    "download_gb": 9.3,
    "ram_gb_needed": 12
  },
  {
    "label": "qwen3:32b",
    "download_gb": 20.0,
    "ram_gb_needed": 24
  }
]

The qwen3:8b build dey 5.2 GB for disk and need about 8 GB RAM to run without swapping. 4 GB VPS no go load qwen3:14b, wey be 9.3 GB before you type even one word into am. The biggest row here, qwen3:32b, need about 24 GB, and for most price lists, na different plan and different monthly bill be that.

Whether e fit enter memory na one question. Speed na another one. For CPU-only VPS, bottleneck na memory bandwidth, no be clock speed. So model wey fit enter memory still fit answer only a few tokens per second. That one okay for job wey run overnight, but e no pleasant for chat. GPU fit increase that figure by roughly one order of magnitude, and e go increase your bill too. So decide with measurement: benchmark the VPS with the workload you plan to run and read when a GPU VPS earns its price. To make weights run at all, start with Ollama on your own VPS.

Assistant na LLM plus chat surface

Assistant na product layer wey dey around model. ChatGPT and Claude na assistants: model, chat window, account, saved conversations, rate limit. Almost none of these dey run for hardware wey you control. Na why hosted assistant dey cost subscription and zero RAM.

The self-hosted version na front end like Open WebUI wey point to local Ollama or hosted API. The front end na small software. Expect about 1 GB resident for the chat surface, on top of wetin the model need. But e need public URL and certificate, while bare model no need am, because you want reach am from phone: issue the certificate with Certbot and Nginx, or terminate TLS at Traefik in front of several apps. If you never choose front end yet, compare Open WebUI alternatives.

Assistant dey answer. E no dey act. When e write shell command, person go read the command and decide whether to paste am. That person na safety layer, and agent na the thing wey remove am.

Agent dey add tools and one loop

Agent na assistant wey fit call functions, then read the results. Two parts dey do the work. The first one na tool: description of function wey model fit request, plus your code wey actually run am. The second one na loop: your program call model, model ask for tool, your program run am, add the output to the conversation, then call model again. This one repeat until model talk say e finish, or limit stop am.

Loop na ordinary code, and basic one fit dey under one hundred lines. Wetin make am agent be say the tools get real credentials, so the loop fit change something outside itself. Na this fact drive every hosting decision wey follow. Agent skills and MCP (model context protocol) servers na two ways to give agent more tools without rewriting the loop.

• E outlive your session. Chat end when you close the tab. Agent run fit take twenty minutes and e suppose survive when your laptop go sleep, so e belong for box wey dey on, started by systemd service or timer wey go bring am back after reboot.
• E hold secrets. API key, SSH key, database password. Anything wey agent fit read, hostile instruction wey hide inside the input fit make am use, so keep the secrets out of the agent's reach.
• The cost dey grow with the loop, not with your question. Every step send the whole conversation again as input, so ten-step run go pay for that transcript ten times. Na why input tokens dominate an agent bill and why you want a hard cap on what one run can spend.
• E fit make mistake wey go write something. Wrong answer for chat only cost you another reread. Wrong delete inside loop fit cost you the directory. Run am as user with the least privilege wey still let am work, and for coding agents, sandbox am before you give am your repository.

Wetin each layer need from the box

[
  {
    "label": "LLM (weights you host)",
    "ram_gb": 8,
    "gpu": "helps a lot",
    "public_url": "no",
    "credentials": "none"
  },
  {
    "label": "Assistant (chat surface)",
    "ram_gb": 1,
    "gpu": "no",
    "public_url": "yes",
    "credentials": "one login"
  },
  {
    "label": "Agent (tools and a loop)",
    "ram_gb": 2,
    "gpu": "no",
    "public_url": "only for webhooks",
    "credentials": "several"
  }
]

Read the RAM column well, because e no include the model. Chat front end and agent runtime na both small programs. If the agent dey call hosted model, 2 GB of RAM fit run am, and cheap plan na correct answer, no be compromise. If you put the weights for the same box, the 8 GB model line go pass everything else by far.

The other columns matter pass wetin people dey expect. Na only model layer dey faster with GPU. Na only assistant layer need public URL as normal requirement, because browser need reach am; agent need one only when something outside need call inside, like webhook. And agent dey hold several credentials, and na this be the real difference between am and chat window. Chat window fit dey wrong. Agent fit dey wrong and still act based on the wrong answer.

You need GPU to run AI agent?

No, except say you dey host the model weights for the same machine. The agent loop na HTTP requests, JSON parsing, and subprocess calls. CPU dey almost idle while e dey wait for network. The GPU question really na question about the LLM layer.

So divide the decision. If the text no fit comot from your machine, pay for enough memory to hold model, and, if you want usable speed, pay for GPU to run am. If na only automation you want, rent the model by token and use the money for uptime and backups instead. Most self-hosted agents for 2026 dey call hosted model, and na why dem cheaper to run.

Agent fit dey self-host?

Yes, and na the agent layer make most sense to self-host, because na the loop hold your data and credentials. Small VPS wey get 2 GB RAM, service manager, and outbound network access fit run real agent. Choose the build-your-own route on a VPS if you want control the loop yourself, or deploy one of the ready-made self-hosted agents if you prefer start with something wey don ready.

To self-host the assistant easy: na one container and one certificate. To self-host the model na the expensive part, and na there many people dey give up after dem watch tokens dey crawl out from CPU. Self-host the weights when the data no fit leave the box, or when your usage volume make per-token pricing too costly. Otherwise, make the agent call an API and keep the important parts local.

ChatGPT na AI agent?

Chat product go become agent immediately e fit call tool and act on the result without asking you first. By this test, hosted assistants wey get browsing, code execution, or connectors na agents. The difference for you na where the loop dey run and whose credentials e dey use. For hosted product, vendor own both. For your own server, you own both, together with responsibility for anything wey the loop do for 3 in the morning.

Reactive, planning na multi-agent

Roundup dem like list seven types of agent. Plenty of those types na just marketing. Two differences dey change the code wey you write, and one dey change the bill. Reactive agent dey call tool, read the answer, then reply. Planning agent first dey write plan, then e work through am. This one dey handle long jobs better, but e dey cost more tokens because dem dey send the plan again for every step. Multi-agent setup let one agent start other agents. E dey multiply token spending and ways wey things fit fail at the same time. So e only make sense when the sub-jobs really independent, like searching four sources at once. Start with reactive. Add planning when runs begin long. Use multi-agent last. For the wider map, see wetin worth learning about AI agents for 2026.

How to know which layer you dey actually run

For server, ask which processes dey hold the memory.

free -h
ps -eo rss,comm --sort=-rss | head -5

If top line na ollama or llama-server wey dey hold several gigabytes of RSS (resident set size, na the memory wey process dey really occupy), na you dey host the model. If nothing pass a few hundred megabytes and your API bill dey continue grow, na you dey host an agent or assistant, while you dey rent the model. If that list empty because everything dey happen for browser tab, na you be customer of an assistant. That one dey okay until you need software wey go act on your behalf.

Wich one you wan run?

• To keep the text private, run the model: host an LLM by yourself with Ollama, then compare runtimes with Ollama against vLLM once one user don turn ten.
• To own the loop and the tools, build the agent: build your own AI agent for VPS.
• To get something wey dey work this evening, deploy one finished agent from self-hosted agents wey worth running.
• If none of these get server under am yet, start from wetin VPS really gives you.

FAQ

AI agent na just LLM wey get extra steps?

The extra steps na the actual product. LLM dey turn text to text, and nothing more. Agent surround am with tools wey e fit call and loop wey continue to call dem. Those tools carry credentials, so output fit change file, database, or live service. Na why agent need machine wey dey on, service manager, and secrets policy. LLM only need enough memory to hold im weights while e dey answer.

I need GPU to run AI agent?

No be for the agent itself. The loop na HTTP requests, JSON handling, and subprocess calls. Any CPU fit manage these things while e dey wait for network. You only need GPU when you host the model weights by yourself and want more than few tokens per second from dem. Agent wey dey call hosted model fit run well for small VPS without any GPU.

How much RAM VPS need for AI agent?

About 2 GB when agent dey call hosted model, because na runtime, im dependencies, and small local database e dey hold. Add the model on top if you host the weights: qwen3:8b alone need around 8 GB. So, all-in-one box go start from that level and increase based on the model wey you choose.

I fit self-host AI assistant and keep my conversations private?

Yes, but one caveat dey decide everything. Self-hosted front end like Open WebUI dey keep accounts and history for your server. The conversations go remain private only if the model behind am dey local too. If you point the same front end to hosted API, the text still dey leave your box for every message. So you keep the history, but you no keep the privacy.

Wetin be the difference between AI agent and chatbot?

Chatbot dey reply and stop. Agent dey decide wetin to do next, call tool, read the result, and decide again until e finish the job or limit stop am. The practical test be this: if the software fit change something without human pressing button between the answer and the action, na agent be that. E need the hosting and guardrails wey come with am.