What Are AI Agents, Really? Four Types and Which Ones Run Locally (June 2026)
“AI agent" is a phrase I have been seeing everywhere lately. But “agent" covers a lot of ground, and what is actually inside varies a great deal from tool to tool. I wanted to know what the differences really were, so I went through the AI agent apps and frameworks people were talking about as of June 2026, sorted by type.
Broadly, it helps to split them into four: coding agents (Claude Code, Codex and so on), general-purpose ones that operate your PC or browser (Manus, Operator), frameworks for building your own agents (LangGraph, CrewAI), and local ones that run on your own machine (Hermes Agent, ollama integrations).
This is a research article compiled from official sources and reviews in June 2026 (checked 16 June 2026). Pricing, performance and availability change, so please check current information. Sources are listed at the end.
What is an AI agent, exactly?
Unlike a chat AI, where you ask and it answers, an AI agent takes an instruction, breaks it into steps on its own, uses tools (file operations, running commands, web search) and sees the job through. It does not just answer — it does.
The more you hand over, though, the greater the risk of it doing the wrong thing, and the more compute you need — meaning hardware, if you are running it locally.
The four types, roughly
[kimono_cost_table title="The four main types of AI agent (as sorted in June 2026)" note="The examples are what people were talking about at the time of research. Sources at the end."]
Type | What it does | Character | Examples | Where it runs
Coding | Reads, writes, fixes and tests code | Automates development work. Lives in your IDE or terminal | Claude Code / Codex / Cursor / Cline | Mostly cloud (some local)
General / PC control | Operates your screen, browser and PC on your behalf | Research, bookings, making documents — PC work in general | Manus / Operator / Devin | Mostly cloud
Agent frameworks | A base for building your own agents | For developers. Coordinates multiple agents | LangGraph / CrewAI / Dify | Self-hosted (cloud or local)
Local / self-run | All of the above on your own PC with a local model | Free, private, and only as good as your hardware | Hermes Agent / Cline+ollama | Local
[/kimono_cost_table]
Three axes tell them apart: what it operates (code / screen / arbitrary tools), where it runs (cloud / your own machine), and who it is for (general users / developers).
1. Coding agents
The kind that reads your code, writes it, and runs the tests. Reviews in 2026 report that the field has become closely matched.
On benchmarks — shared tests that compare tools on the same tasks; here, Terminal-Bench 2.1 — Codex CLI (GPT-5.5) leads at about 83%, with Claude Code (Opus 4.8) a few points behind, while on SWE-bench Pro there are reports of Claude Code leading. Rankings swap depending on the task and the metric, and the gap between the leaders is reported to be narrowing.
- Claude Code / Codex CLI / Gemini CLI: CLI tools (you type commands into a terminal) driving each vendor’s cloud models. Strong at deep automation. Billed by API usage or subscription. For people who want the newest models and want to hand over difficult work.
- Cursor: A VS Code–based AI editor with its own models, completing the work inside the editor. Mostly cloud.
- Windsurf: Described as a hand-off model — start in the editor, pass the work to a cloud autonomous agent (Devin), get the result back.
- OpenCode / Cline / Aider / Goose: Open source, and they work with local models (ollama, LM Studio). OpenCode has one of the highest GitHub star counts; Cline lives inside your editor; Aider is a terminal tool built around git. For people who want free, self-run setups.
2. General-purpose / computer-control agents
Not limited to code — these operate the browser or the PC itself on your behalf. In 2026, Devin, Operator and Manus are frequently introduced as the “big three".
- Manus: A general-purpose autonomous agent that goes from a single instruction to browser operation, running code and producing documents. Its March 2026 desktop version, “My Computer", is said to operate your own PC — reading and writing files, running commands, driving applications, with a permission gate per command. Light work happens locally, heavy inference in the cloud. Note that an acquisition by Meta (reported at around $2bn) was later reported to have been blocked and unwound by Chinese authorities in April 2026.
- OpenAI Operator: ChatGPT’s autonomous web agent. Described as a computer-control type that handles research and bookings with human checkpoints along the way.
- Devin (Cognition): Known as an autonomous software engineer, also presented in combination with the Windsurf editor.
The upside is handing over whole PC tasks; the downsides are the risk of wrong actions and the cost, since these are mostly cloud services.
3. Agent frameworks
Not finished applications, but the foundations for building your own agent. For developers.
- LangGraph: A base for building complex, stateful agents. Enterprise adoption is growing and monthly downloads are reported to be very high.
- CrewAI: Its selling point is assembling multiple role-based agents quickly.
- Dify: Open source, low-code agent and app building. High GitHub star count.
- OpenAI Agents SDK / AutoGen / Smolagents and others, from various vendors and open-source projects. Most support multiple model providers and can connect to a local model server. (AutoGPT was the pioneer here but is described as drifting out of the mainstream.)
4. Running it locally, on your own machine
This is the part closest to what this blog is about. Running an agent on your own PC with a local model, without relying on a cloud API, has become a real option.
- Hermes Agent (Nous Research, released February 2026): An open-source autonomous agent framework. Model-agnostic, with persistent memory, and using a simple function-calling format (XML + JSON) that is said to sit well with local models and general frameworks.
- ollama + tool calling: Getting a local model to call functions, as the base for your own agent.
- Coding agents with local support (Cline / OpenCode / Aider / Goose) pointed at a local model — a common setup.
The upsides: it is free to run, your code and information never leave the machine, and there are no usage caps. The downside: quality and speed depend entirely on the hardware in front of you.
Running it on your own hardware
For local agents, the deciding factor is whether the model you want fits in your memory (VRAM or RAM). The RAM guidance in various guides runs to 16–32GB for small, quantized models (compressed to be lighter, at a small cost in accuracy), 32–64GB for mid-size coding models, and 64GB or more for large ones.
For which sizes are actually practical, my measurements of several models on my own machines are worth a look — the short version is that a model with a small active portion that fits on the GPU is the fast one.
Choosing: cloud or your own machine
- Newest and highest performance, handing over development → a cloud coding agent (Claude Code, Codex).
- Handing over the PC work itself → a general-purpose control type (Manus, Operator).
- Designing your own agent → a framework (LangGraph, CrewAI).
- Free, privacy-first, everything on your own machine → local (Hermes Agent, Cline + ollama). But you need the hardware.
Wrap-up
AI agents come in these four shapes — coding, PC control, frameworks, and local — each good at different things. As of June 2026 the cloud frontier (Codex, Claude Code, Manus) leads on performance, while self-run local options (Hermes Agent, local-capable agents plus ollama) have become practical. Cloud for the hard parts, local for volume, is a realistic way to split the work.
Sources
Official sites for the main tools mentioned (checked 16 June 2026). Pricing and availability change; please check the official sources for the current position.