AI Pro League
Season Zero
The agent products and frameworks that turn language models into workers: what each one does, what it costs, and how it is really adopted — every figure sourced and dated.
An AI agent is a system built around a language model that does not just answer — it plans a multi-step job and carries it out: running code, browsing the web, editing files, calling APIs and checking its own results, usually with a human reviewing the outcome. A chatbot replies; an agent delivers. The practical differences are where the agent runs (a vendor cloud sandbox, your editor, or your own server), how much supervision it takes, and what it costs per finished task.
This section tracks the two halves of the market. Agent products are finished workers you hire by the month — coding agents, general-purpose cloud agents and self-hosted runtimes. Agent frameworks are the open-source libraries developers build their own agents with. Every price, star count and status below carries its source and the date it was checked; where no honest public metric exists, we say so instead of inventing a score.
| # | Project | Category | GitHub stars | Status |
|---|---|---|---|---|
| 1 | OpenClaw | Self-hosted agents | ≈388K | active |
| 2 | Hermes Agent | Self-hosted agents | ≈249K | active |
| 3 | OpenAI Codex | Coding agents | ≈62.0K | active |
| 4 | AutoGen | Agent frameworks | ≈61.2K | maintenance |
| 5 | CrewAI | Agent frameworks | ≈59.2K | active |
Full leaderboard & methodology →
Agents whose job is software: they take tickets and tasks, edit code, run tests and open pull requests.
Cognition
Devin is Cognition's autonomous AI software engineer: a cloud-hosted agent that takes a task description, plans and executes it end-to-end in its own sandbox (shell, code editor, browser), and delivers the result as a pull request.
Subscription + usage credits
Full profile →Anthropic
Claude Code is Anthropic's agentic coding product: the same agent core drives a terminal CLI, IDE integrations, a desktop app and a web interface.
Bundled subscription / API usage
Full profile →OpenAI
Codex is OpenAI's coding agent system, relaunched in 2025 across four surfaces: an open-source Rust CLI (about 62,000 GitHub stars), an IDE extension, cloud tasks inside ChatGPT, and an SDK.
Bundled subscription / API usage
Full profile →Anysphere
Cursor is Anysphere's AI code editor: a VS Code-derived IDE where agent mode, background agents and best-in-class tab completion sit on top of a multi-model catalog.
Subscription + usage credits
Full profile →GitHub (Microsoft)
GitHub Copilot is the incumbent AI coding assistant, built into VS Code, JetBrains, Neovim and the GitHub platform itself.
Per-seat subscription
Full profile →Cloud agents that take open-ended goals — research, documents, data, web apps — and deliver finished work.
Manus
Manus is a general-purpose cloud agent: give it a goal and it researches, analyses data, builds spreadsheets and slides, and — since the 1.5 release — generates full web applications with backend and database.
Subscription + credits
Full profile →Open-source agents that run on your own machines, with your own model keys and your data staying home.
OpenClaw (open-source project)
OpenClaw is the open-source personal agent that went viral in early 2026: a self-hosted runtime (created by Peter Steinberger, first released in November 2025 as Warelay, then renamed Clawdbot and Moltbot before settling on OpenClaw) that lives on your own machine, keeps memory and skills as files on your disk, and is controlled through WhatsApp, Telegram, Slack, Discord, Signal, iMessage and 20+ other channels.
Free / open source
Full profile →Nous Research
Hermes Agent is Nous Research's self-improving open-source agent, launched February 25, 2026 under the MIT license.
Free / open source
Full profile →The libraries developers build agents with: orchestration, memory, tools and guardrails. You bring the model.
LangChain
LangGraph is LangChain's low-level orchestration library for stateful, long-running agents, modelled as directed graphs with built-in checkpointing — a graph can fail, resume exactly where it left off, and replay earlier states.
Free / open source (+ paid platform)
Full profile →CrewAI
CrewAI is the role-based multi-agent framework: instead of drawing graphs, you describe a team — each agent gets a role, a goal and a backstory — and the framework runs the collaboration.
Free / open source (+ paid platform)
Full profile →Microsoft
AutoGen, Microsoft Research's conversation-driven multi-agent framework, pioneered the multi-agent patterns most frameworks now copy — and it is the cautionary row in this section: with about 61,000 GitHub stars it is the most-starred framework here, yet Microsoft moved it to maintenance mode (bug fixes and security patches only; last Python release v0.7.5 on September 30, 2025) and its README directs new users to Microsoft Agent Framework.
Free / open source
Full profile →Microsoft
Microsoft Agent Framework is the convergence of AutoGen and Semantic Kernel: announced October 1, 2025, version 1.0 shipped in early April 2026 for Python and .NET under MIT.
Free / open source
Full profile →OpenAI
The OpenAI Agents SDK launched March 11, 2025 alongside the Responses API, as the production evolution of OpenAI's experimental Swarm project.
Free / open source
Full profile →LlamaIndex
LlamaIndex is the leading framework for building LLM applications over your own data — indexing, retrieval and RAG first, with agent workflows on top.
Free / open source (+ paid cloud)
Full profile →AWS
Strands Agents is AWS's open-source agent SDK, launched in preview in May 2025 and at 1.0 in 2026, taking the opposite bet from explicit-graph frameworks: the model drives the loop.
Free / open source
Full profile →How this section stays honest: prices are quoted from the maker's published tiers as reported by the cited sources on the date shown; GitHub stars are dated snapshots, not live counts; a framework in maintenance mode is labelled, not quietly dropped. Agent products share no public benchmark, so there is deliberately no composite "Agent Score" — the leaderboard ranks only what can be measured in public.