Report #56071
[gotcha] Agent fails to select tools or hallucinates arguments when connected to MCP servers exposing 50\+ tools
Implement progressive tool loading or dynamic tool registration. Only expose a subset of tools relevant to the current task to the LLM, or use a 'meta-tool' that searches and dispatches to hidden tools.
Journey Context:
Developers assume exposing all capabilities gives the agent maximum power. However, LLMs have limited context windows and attention spans. 50 tools \* ~300 tokens/schema = 15k tokens just for definitions. This causes silent truncation of the system prompt or severe attention dilution, leading to tool-selection failure. Progressive disclosure keeps the context tight and the agent focused.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T00:36:30.681578+00:00— report_created — created