An AI Agent (an AI assistant that can call external tools) given 50 tools will often perform worst if you dump them all into the model — that's the core finding of a recent AI pharma paper. The paper's solution: first match the user's question to a "skill template" (similar to a standard operating procedure, or SOP), pick the must-have tools from that template, then supplement with other tools using one of three modes.

What this is

The three supplementation modes are:

  • LLM mode: Let the model read tool descriptions and pick on its own — saves tokens (models charge by input text volume) but risks wrong picks
  • Embedding mode: Convert both questions and tool descriptions into numerical vectors (a method that lets machines understand semantic distance), rank by similarity — similar to RAG (retrieval-augmented generation, where AI retrieves documents before answering)
  • All mode: Stuff everything into the context window (the total text the AI can "see" at once). Simplest, but the more tools, the more the model's "selection attention" gets diluted — and accuracy drops

The paper also designates several "core tools" (like running Python, querying PubMed) that stay in regardless of the question — they're never filtered out.

Industry view

Mainstream frameworks like LangChain and AutoGPT default to all mode — fast to build, broadly applicable. The paper's authors argue that in high-stakes domains like healthcare and law, pre-filtering is non-negotiable — otherwise, models facing a pile of tools tend to "grab the wrong medicine."

But there's clear pushback. An enterprise AI architect told us directly: in consumer scenarios, adding a matching layer means extra maintenance cost and latency; embedding mode often produces "looks-like-but-isn't" matches on ambiguous questions, forcing errors in. Translation: the paper's solution fits specialized scenarios but may be over-engineering for general chatbots.

Impact on regular people

For enterprise IT: When evaluating Agent vendors, ask one more question: "How do they handle tool selection beyond 30 tools?" — the answer reveals real cost and accuracy gaps.

For working professionals: When collaborating with AI, proactively telling it "follow this process" yields noticeably higher efficiency than throwing complex questions at it cold.

For consumers: Smart assistants that "remember your habits" (e.g., booking flights based on your travel template) will leave behind products that require you to repeat yourself every time.