What is an agentic assistant, and why build your own

You have probably chatted with ChatGPT, asked it questions, and had it draft an email. But you have also hit the wall: a chatbot can only talk back. In other words, it cannot rename a file, check a folder, schedule a job, or fix a broken line of code on its own. That gap is exactly what an agentic assistant exists to close. This article explains what an agentic assistant actually is, and how it differs from a plain chatbot. It also covers why more people are running their own on an ordinary Windows 11 PC instead of renting one from a company.

A chatbot answers; an agent does

The real difference between a chatbot and an agent is not how smart either one is — it is what happens after the reply. A chatbot like ChatGPT or Claude generates text inside a chat window and stops there. It cannot touch your machine.

An agentic assistant plans and carries out multi-step tasks by calling tools — running terminal commands, editing files, driving a browser — with some autonomy. It also keeps memory across sessions, so it remembers context between conversations. Anthropic, one of the companies behind these systems, draws the line in its guide to building effective agents. A workflow is a language model plus tools orchestrated by code written in advance. An agent, in contrast, directs its own process and chooses its own tools as it goes.

A concrete example makes it obvious. Ask a chatbot “why is my code broken?” and it explains what is probably wrong. Ask Claude Code the same question and it reads your repository, edits the file, runs the tests, and commits the fix. A chatbot can draft an email for you; Hermes Agent can be scheduled on a timer to build a whole report and deliver it to your Telegram every morning. Same underlying model — a completely different level of usefulness.

What an open-source agent can actually do

The easiest way to understand the category is to look at a real one. Hermes Agent, from Nous Research and released under the MIT license, runs as a command-line tool or a desktop app with a concrete set of abilities:

• terminal execution — run commands on your machine
• file reading, editing, and patching
• browser automation
• a cron scheduler for recurring tasks
• persistent memory across sessions
• reusable “skills” — small procedures it writes and saves from its own experience
• delegation to sub-agents for parallel work
• MCP integration to reach external tools and data

In short, MCP — the Model Context Protocol, an open standard Anthropic introduced in November 2024 — is the piece worth knowing. Think of it as a “USB port” for agents: one connector that lets a single agent plug into many different tools and data sources.

The closed, commercial equivalents — Claude Code and OpenAI Codex — do the same terminal, editing, and browser work. They are excellent, but they are products you rent rather than software you own.

Why run your own instead of renting

Four reasons keep coming up.

Privacy and control. A chatbot routes every prompt through one vendor’s cloud. A self-run agent lets you pick any model provider — or run a model fully locally with tools like Ollama or LM Studio — with, in Hermes’ own words, “no lock-in.”

Cost. Paying per token with a cheap model can undercut a flat monthly subscription, especially if you automate small jobs.

Customization. Skills, MCP servers, plugins, and personality files let you shape the assistant to your workflow.

Learning. Running one yourself teaches you the actual stack underneath — terminals, models, and providers — instead of hiding it behind a polished app.

The Windows 11 reality

Importantly, you do not need a server or a powerful machine. Hermes Agent and Claude Code both install on native Windows through a PowerShell one-liner, and Hermes also ships a desktop installer. The installers bundle their own dependencies (Python, Node.js, ripgrep, and ffmpeg), so there is little to set up by hand.

Here is the key point: no GPU is needed for API-based agents. The model runs in the vendor’s cloud, so an ordinary laptop is plenty — Hermes’ README notes it runs “on a $5 VPS.” Hardware demands only appear if you choose to run open-weight models entirely on your own machine. (One naming trap to avoid: “Hermes Agent” the tool is not the same thing as “Hermes 3,” which is a model.)

The landscape and an honest look at pricing

The open-source and commercial landscape

On the open-source side, your options include Hermes Agent, OpenHands, Aider, Cline, Goose, and OpenCode. On the commercial side sit Claude Code and OpenAI Codex — both climbing fast. As of August 2026 they sit around 142,000 and 111,000 GitHub stars respectively. Star counts only grow, so read these as a dated checkpoint, not a live total.

Two pricing models

Pricing comes in two shapes, and it helps to know which you’re on:

• Pay-as-you-go (per token). You are billed per token — a token is a small chunk of text, roughly a word or a bit of one. Input (what you send) and output (what it generates) are charged separately, quoted per one million tokens. Your bill scales with how much you use.
• Subscription. A fixed monthly fee, with the token-metering hidden under the hood. Claude Code inside a Claude Pro or Max plan works this way: a predictable cost whether you use it lightly or heavily.

Three tiers, by role

However, the numbers really only vary on the pay-as-you-go side, and they move week to week. So it helps to think in tiers by role rather than memorise figures. Two things stay true even as the figures shift: output runs roughly 5× the price of input, and the spread from cheapest to most expensive is more than 100×.

• Budget (open-weight). The cheapest tier — open models like DeepSeek or MiniMax. Best for high-volume, straightforward work where you do not need a flagship.
• Mid (workhorse). The balanced daily driver — Claude Sonnet 5, GPT-5.6 Terra. Roughly an order of magnitude above budget, and enough for most real tasks.
• Frontier (flagship). The heavyweights for the hardest reasoning, agentic work, and complex code — Claude Opus 5, GPT-5.6 Sol. Tens of times the cost of the budget tier.

Why the numbers move

For example, a couple of quirks come and go. Some providers run cheaper off-peak windows and pricier peak hours. Temporary discounts and free preview models appear and vanish too. So treat every figure as a snapshot — for today’s number, check the provider’s own pricing page — Claude’s, for instance.

Honest limitations

Agentic assistants are not a magic upgrade to everything. Anthropic is blunt about the trade-off: these systems “trade latency and cost for task performance,” and often a single, well-written prompt to a normal model is all you actually need.

For example, when they fail, they tend to fail semantically — a hallucinated tool call, a misread instruction, a corrupted bit of memory. It is rarely a clean error code you can google. There is a real learning curve: terminals, configuration files, and model providers are part of the deal. And it is the wrong tool for one-off questions, or for anything you cannot review after the agent acts.

Where this series goes next

This article is the pillar of a new series: Building your own agentic AI assistant. It is written for people who know ChatGPT but have never run an agent. Over the coming articles we will walk through it step by step. You will install your first agent on Windows 11, give it tools and MCP connections, teach it skills and memory, schedule automations, and choose a model provider that fits your budget.

If you have only ever used a chatbot, an agentic assistant is the natural next step — the same language model, but one that can actually reach out and do things on your machine. Start with the difference we just covered, and when you are ready, the rest of this series will get you running your own on the PC you already own. You can follow every part as it is published from the series overview page.