What is an AI agent? Types, frameworks and real examples (October 2026)
Every product in 2026 calls itself an agent. Most are chatbots with a button. Here is a working definition, the four types that actually matter, the frameworks worth your time, and real plugins you can install tonight โ from the team that indexes 350+ of them daily.
The working definition
An agent = a model + tools + a loop. The model decides what to do, the tools let it act (edit files, fetch pages, call APIs), and the loop lets it check its own work and continue. Take away the loop and you have a chatbot; take away the tools and you have a text generator.
That definition explains why plugins matter so much: a harness like dsh or Codex is the loop, and plugins are the tools. More tools, more things the agent can actually do.
The four types that matter
Work in your repos: edit, run, verify. dsh (free, plugin-extensible), Codex (commercial). The most mature category.
Browse, extract, act on pages. Jev is the current speed/cost leader for high-volume extraction.
Tiny fast models (Laya, Kev) making typed yes/no or scored choices โ the quiet workhorse inside agent pipelines.
Magpie-style model routers: Codex on DeepSeek, Claude Code on Kimi โ pick the model per task without re-wiring.
Real plugins you can install tonight
- dsh-our-free-model โ frontier models (DeepSeek V4.1 Flash, Kimi K3) in dsh with zero API keys. 40k+ stars.
- dsh plugin directory โ 150 indexed plugins, refreshed daily.
- Codex plugin directory โ 198 plugins for the OpenAI Codex ecosystem.
- Laya vs Jev comparison โ picking between open decision models and managed web agents.
Every entry in our directory is checked and re-indexed daily โ no dead projects, no empty READMEs.
FAQ
What is an AI agent in simple terms?
An AI agent is software where a language model does not just answer โ it acts. It plans a few steps, calls tools (search, code, files, APIs), checks the result, and keeps going until the task is done. ChatGPT answering a question is a chat; an agent that reads your codebase, edits five files and runs the tests is an agent.
What are the main types of AI agents in 2026?
Four families dominate: (1) coding agents โ Claude Code, Codex, dsh โ that work inside your repos; (2) web agents โ like Jev โ that browse and extract; (3) decision models โ Laya, Kev โ tiny fast models that make yes/no or scored choices inside a bigger pipeline; (4) orchestrators โ like Magpie โ that route work across models.
Which agent framework should I start with?
If you write code, start with a coding agent harness (dsh is free and plugin-extensible; Codex is the commercial option) and add plugins from a directory rather than building from scratch. If your problem is high-volume simple decisions, look at decision models instead โ an agent is overkill for a yes/no.
Do agents need API keys and subscriptions?
Not always. Routes like the dsh-our-free-model plugin expose frontier models with no API key, and open-weight decision models run locally. For production work, bring your own keys โ free routes are for trying things out.