Sixteen frameworks, sixteen folders, and no lockfile

Sixteen top-level folders, one per framework

The fastest way to read this repository is to read its directory listing. Every entry is a framework, and there are fifteen of them plus one folder that is not: ag2, agno, autogen, claude-agents-sdk, crewai, deepagents, google-adk, langchain, langgraph, llama-index, microsoft-agent-framework, openai-agents-sdk, pydantic-ai, smolagents and strands-agents-sdk.

That is a wider net than most comparisons of this kind take. It spans general agent runtimes from several vendors, a hosted vendor SDK wrapped as a Python package, at least one orchestration library that is not itself an agent, and one agent project for AWS. The primary language is Python throughout, so the language choice does no filtering here.

The structural consequence is that there is no shared package to install. Each framework owns its own folder, its own dependency expectations and its own idioms, and the repository itself sits above all of them as a place to read and compare rather than a library you depend on. Whether any of the folders shares a common environment is not something the top-level listing answers, which is the first thing to check if you plan to run more than one example.

The version table is a snapshot, not a lockfile

The README leads with a table whose columns are framework, version, documentation and repository. The versions visible in it are ag2 at 0.13.4, agno at 3.0.9, Autogen at 0.7.5, Claude Agent SDK at 0.2.130, CrewAI at 1.15.21 and Deep Agents at 0.7.13.

Read those numbers as a photograph, not a constraint. They record what was current when the table was last edited, and nothing in the top-level listing ties them to a requirements file, a lockfile or a container. That matters because agent frameworks move fast and change import paths between minor versions, so a table entry tells you which release the examples were written against and nothing more.

The two columns beside the version are the load-bearing part of the table. Each row links the project's documentation and its source repository, which turns the table into a set of hand-checked entry points rather than a summary you have to trust. There are also links for reporting a bug and requesting a feature, pre-labelled with bug and enhancement.

Six documentation hosts, no single one in charge

Following the documentation column shows how fragmented the landscape is. The ag2 row points at docs.ag2.ai. The agno row points at docs.agno.com. Autogen is hosted by Microsoft at microsoft.github.io. The Claude Agent SDK row points at platform.claude.com. CrewAI has docs.crewai.com. Deep Agents is documented under LangChain's own domain at docs.langchain.com.

Three hosting patterns in six rows. A vendor runs its own docs, in the case of ag2, agno, CrewAI and the Claude SDK. A code-hosting project page serves the docs, as with Microsoft's Autogen site. And a large ecosystem vendor documents someone else's framework inside its own domain, which is Deep Agents living under LangChain's documentation tree.

That last pattern is the one worth pausing on when reading the table. Whether a framework is documented under the vendor's own name or inside another company's docs is a decent signal of how closely the two projects are tied in practice, and the table makes that visible without saying it in prose.

Logos ship twice so the table renders in dark mode

The table is built out of image files in a res/ directory, and there is more plumbing here than the visible result suggests. Every framework logo has a light and a dark variant, ag2-dark.svg and ag2.svg, and each one is wrapped in a picture element with a source for prefers-color-scheme: dark and another for the light case. The documentation and repository icons follow the same pattern, with book-open-dark.svg, book-open.svg, github-icon-dark.svg and github-icon.svg.

So the res/ folder is not decoration. It is a set of paired assets wired into media queries, which is why the README renders with vendor branding on both colour schemes instead of a washed-out logo on a dark page.

There is a second consequence. Some rows carry their name as text in a strong element and some rely on the image alone, which is why the first entry in the table is recognisable mainly by its filename. A reader scanning quickly gets the friendly vendor marks and misses the machine-readable ones, which is the kind of small inconsistency that shows the table was maintained by hand over time rather than generated.

Families get separate rows, and Microsoft gets two

Several entries exist because a single vendor or ecosystem has more than one thing worth comparing, and the listing treats them as separate subjects. LangChain and LangGraph have their own folders, as do llama-index and the Microsoft Agent Framework alongside autogen. pydantic-ai, smolagents, google-adk, openai-agents-sdk and strands-agents-sdk each get one.

The Microsoft pair is the clearest case. autogen is listed as a row in the version table with its own version and its own links, and microsoft-agent-framework has its own folder as well. Two frameworks from one organisation, treated as two entries, which is the correct treatment for a reader trying to choose between them.

The same logic separates langchain from langgraph: one is the library most examples import, the other is the orchestration layer for graph-shaped agents, and a comparison that merged them would hide the choice the reader is actually making. This repository does not merge them, which is the clearest signal of its editorial approach: where the ecosystem splits, the folder splits too.

study-agents-differences is the only folder that is not a framework

Among the framework folders there is one entry whose name describes an activity rather than a project: study-agents-differences.

Its presence tells you what the repository thinks its job is. A collection of examples could stop at sixteen folders and let the reader diff them unaided. This one adds a folder whose purpose is the comparison itself, which matches how the repository describes itself as a place to learn, experiment with and compare open-source agent frameworks.

The description also claims a range from the basics to production-ready setups, which is the one claim in the visible material that the directory listing cannot corroborate. What the listing does corroborate is the breadth, the per-project isolation and the willingness to keep a dedicated study folder. Whether the study material goes from a first agent to a deployment is something you only learn by opening that folder.

No install command, no license, no releases

Three signals are missing, and they are the ones that matter for anything beyond reading.

There is no install command anywhere in the visible documentation. No clone step, no dependency install, no run command for any of the sixteen frameworks. The table hands you documentation and repository links instead, so getting any example running means leaving this repository and following one of those links, which makes it an index rather than an environment.

There is also no license. None appears among the top-level entries, and the license field is recorded as unknown. For a repository that redistributes nothing but examples, that matters less than it would for a library, but it does mean the examples themselves carry no stated terms.

Finally, there are no GitHub releases, so there is no tagged version of the comparison to point at, and the default branch is main. The last push landed on 2 October 2026, which means the version table is being maintained as the frameworks move. Fresh, then, but fresh is not the same as pinned: check each framework's current release before you follow its row.

Editorial conclusion

Use it as an index: when you are choosing between two agent frameworks and want their landing pages and repositories in one place, the folder listing and the version table answer that in ten seconds. Do not expect a runnable comparison, because the visible material carries no install command, no pinned environment and no license, and it says nothing about how the examples in each folder are meant to be run. Before you adopt anything it points you at, check the version numbers yourself, since they are a snapshot of one moment rather than a lockfile, and read the license of the framework you actually intend to ship.

Frequently asked questions

Official sources

1. Issues
2. martimfasantos/ai-agents-frameworks on GitHub
3. README