Skill degradation at 50+ tools, is this a fundamental self-hosted agent limitation?
Posted by Academic_Cut3518@reddit | LocalLLaMA | View on Reddit | 5 comments
Tool calls degrade past 50+ skills in self-hosted agents. Tried skill tiering, explicit routing, and prompt normalization, all help marginally.
Is there a documented architectural fix, or is this inherent to how most frameworks handle skill dispatch at scale?
Automatic-Arm8153@reddit
Why on earth are you running 50+ skills or tools anyway..
Are you trying to vibecode the matrix?
DistanceAlert5706@reddit
That's pretty easy nowadays, only Playwright MCP alone has like 26 tools, pair it with jira, GitHub MCP and you are way over 50...
Automatic-Arm8153@reddit
Oh interesting, that makes sense
En-tro-py@reddit
I find it very hard to believe you need 50 tools all the time, so either only enable the tools you need, pay the cache hit to shuffle the setup, but you maintain a workable context for the agent...
OR
Continue to kitchen sink it with tool descriptions clogging up your context. Maybe refine them to be more context efficient and to prevent overlap/confusion...
Can you break your flows down and pass off the context?
2-3 agents configured for ~10-20 tools vs 1 with 50
whenhellfreezes@reddit
You seem to be using tool call and skill interchangeably. While yes it takes a tool call to invoke a skill and yes both skill definitions and tool call definitions take up starting context they are best thought of as different things. So it's probably best if you described your starting problem.