Skip to content

ai · 1 min read

Tool Use: Designing Functions Your Agent Will Actually Call Correctly

LLM tool calls fail silently when functions are poorly designed. Here are the eight rules Bhogar AI follows so models pick the right tool with the right arguments.

BABhogar AI TeamProduct & Engineering

Tool calling is where LLM agents meet the real world. Every external API your agent can invoke is a tool - and every poorly designed tool is a future production incident. We have catalogued the eight rules that move a tool from "model usually picks it wrong" to "model picks it correctly 99% of the time".

Why it matters

Most teams expose their existing REST APIs to the agent verbatim and wonder why the model passes the wrong arguments. LLMs are not psychic - they need names, descriptions and shapes designed for natural-language reasoning, not for backend convenience.

How Bhogar AI approaches it

Bhogar AI ships a tool design linter and a registry that enforces our rules: verb-first names, single-purpose actions, explicit enums, pre-condition descriptions, and example arguments embedded in the schema. Tools registered through Bhogar Buddy are automatically scored for "model-friendliness".

  • Tool registry with auto-generated JSON Schema
  • Linter that flags ambiguous parameter names
  • Per-tool allow-list of agents that may call it
  • Built-in dry-run mode for safe testing
  • OpenTelemetry spans on every tool invocation

What you get

Teams that adopt these rules see tool-call accuracy climb from a typical 70-80% baseline to consistent 95%+ on production traffic, often without changing the underlying model.

See Bhogar on your own data

Book a 45-minute working session. We connect one of your sources, build one agent, run one governed workflow, and review the trace together.