github.com/agents-sh/radkit
radkit.rs/getting-started/
Most of the agent librararies out there just adds a huge prompt, put all the tool definitions and call the LLM on a loop hope and for the best.
This is not a way to build reliable ai agents. A better way to build ai agents is described in 12-factor agents by Dex Horthy here.
https://www.youtube.com/watch?v=8kMaTybvDUw
https://github.com/humanlayer/12-factor-agents
Another limitations of current agentic frameworks is their support for interoperability among agents is very limited.
Google came up with https://a2a-protocol.org/latest/ (now under linux foundation) to address this, but the agent frameworks trying to do as much as possible and keep the end users locked into their "ecosystem" instead of supporting such protocols.
Everyone is so focused on MCP for some reason. (no hate for mcp, radkit supports mcp as well).
We belive in an agentic future where businesses expose ai agents and not mcp servers or rest apis. If you build an agent using radkit, it is a2a-protocol compliant from the get go.
With all of this in mind, we are building `radkit`.
It is still early stages. But we have reached a milestone and wanted to share with the community.