
AI developers are wasting thousands of dollars making the same stupid mistake. Hear me out.
you get an idea for an agent.
you open your editor.
you spend the whole weekend building memory, a browser and an agent loop from zero.
then on Monday you find out 5 open-source projects already did it better.
there's a GitHub repo that maps 300+ AI agent projects across 36 categories. it's the first thing I open now before writing a single line.
and the gold isn't the famous names. it's the stuff nobody reposts:
tired of rewriting prompts by hand?
-> DSPy, from Stanford. you write your agent as code, not as prompts
want your coding agent to understand the code, not just read files?
-> Serena. it finds and edits by meaning, like a real IDE
want an agent that can actually use any website?
-> Stagehand. it pulls data and clicks through sites for you
need a browser your agents can run in without breaking yours?
-> Steel Browser. a full browser sandbox, ready to go
want your agent to remember things long-term without a huge bill?
-> Cognee. persistent memory that runs on small models for free
want your agent to know what you actually did on your computer this week?
-> Screenpipe. it records your work so your agents have real context
need to build a whole agent team that you can test and deploy?
-> Google's ADK. code-first, with evaluation built in
want to see exactly where your agent went wrong?
-> Opik. traces, evals and dashboards for every run
the rule I follow now:
pick the category -> grab 3 options -> check which one still gets commits -> build on the winner.
the "still alive" part saves you. AutoGen is on the list, but Microsoft itself now points new users to its successor, Microsoft Agent Framework. pick the wrong one and you're building on a dead end.
and if you're a builder, the whole list also comes as JSON and YAML you can plug straight into your own tools.
free. MIT licensed. no signup.
the repo + the projects I'd start with in each category are below
kiosa @thegreatest_sv