The tool discovery overhead is something nobody talks about enough. I am running an AI agent system with 30+ automation scripts, and the agent wastes tokens just figuring out which tool to call before doing anything useful.
The RBAC per tool is what caught my attention most. My current setup has a crude allowlist approach where each agent session gets a filtered set of tools, but there is zero audit trail when something goes wrong.
Have you seen good patterns for retroactively debugging why an agent chose a specific tool path? That is where I keep hitting walls. Even with structured logging, the decision chain stays opaque, especially when agents chain 3-4 tool calls and the failure only shows up at the end.
The tool discovery overhead is something nobody talks about enough. I am running an AI agent system with 30+ automation scripts, and the agent wastes tokens just figuring out which tool to call before doing anything useful.
The RBAC per tool is what caught my attention most. My current setup has a crude allowlist approach where each agent session gets a filtered set of tools, but there is zero audit trail when something goes wrong.
Have you seen good patterns for retroactively debugging why an agent chose a specific tool path? That is where I keep hitting walls. Even with structured logging, the decision chain stays opaque, especially when agents chain 3-4 tool calls and the failure only shows up at the end.
This module was created specifically to automate work with LLM, for example.