Million-dollar problem in its own way.

After seeing how thousands of users use Ogcode, we noticed something interesting. Many developers have hundreds of skills and MCP tools installed, and honestly, this creates a problem. The agent has too many things to choose from. Sometimes it picks the wrong tool. Sometimes it ignores the tool that could have solved the problem. And sometimes, it gets confused about what it should use. But there's a bigger problem here. We're also wasting tokens. Even when most of these tools, mcp tools and skills aren't needed for the task at hand, their descriptions and instructions can still take up context. This happens turn after turn, and over a long coding session, the waste can become significant larger - so large that it contribute 77% of total token consumed to solve a given task , all wasted ones. ( all data from our internal traffics ) Now imagine if the agent could figure out exactly which skills and tools it needs for a task, without having to deal with everything else- this is not easy as it sounds. Less wasted tokens, better tool selection, and potentially more accurate results. That's the kind of coding agent we want to build. My engineering team is already working on this, but I'd love to hear from people who enjoy solving difficult engineering problems. How would you approach this? I'm genuinely interested in hearing your ideas, even if they sound unconventional. Someone in this community might have a solution we haven't thought of yet.