Problems we experienced:
* We risked our stateless architecture when we wanted to add an agented feature to our existing system. Current frameworks lack server-client architecture, requiring significant effort to maintain statelessness when adding an agent framework to your application.
* Scaling problem - needed to write Docker configurations as existing frameworks lack official Docker support. Each agent in my application required a separate container (e.g., Twitter page analysis, website scraping, automatic documentation writing, etc.), necessitating individual agent deployment and health checks monitoring.
* Needed LLM calls for simple tasks - both fast and cost-effective solutions. With increased model capabilities, the framework should offer this option. I could handle LLM calls myself, but structured outputs required extra work within task structure.
Due to these problems, we decided to build a dockerized agent framework with server-client architecture. Though server-client architecture slowed development, we observe many benefits for users. We're developing a task-centric approach as we expect agents to complete simple tasks and assist with work.
As tool support is crucial for completing tasks, we built a structure officially supporting MCP servers. Client-server architecture proved beneficial for MCP server stability.
Finally, we prioritized simplicity in the framework, developing a structure where outputs are more easily coded with object responses. We'd be very happy if you could check our repo. I'd love to hear any questions you may have.,
