A custom tool is a function you write and expose to an AI agent so it can perform a specific action in your system. Out-of-the-box, an LLM can only generate text. Custom tools are how you give it the ability to actually do things — query your database, send an email through your provider, create a ticket in your support system, kick off a workflow in your business logic.
The anatomy of a custom tool:
- Name — short, action-oriented (create_invoice, search_customer_records).
- Description — clear sentence explaining what it does, in language the LLM will understand. This is critical; vague descriptions produce wrong calls.
- Argument schema — JSON schema defining the parameters: types, required fields, enums, descriptions of each.
- Implementation — the actual code that runs when the model decides to call it.
- Return shape — structured response the model can parse and reason about.
A well-written custom tool, expressed in pseudocode, has a name like search_orders, a description like "Search customer orders by date range, status or product. Returns up to 50 matching orders.", and a parameters schema with customer_id (required string), days_back (integer, default 30, max 365), and status (string enum of pending, shipped, delivered, returned). That structure — short clear name, sentence-long description, tightly typed parameters with sensible defaults — is the entire art of designing tools an LLM can use reliably.
The principles that produce reliable custom tools in 2026:
- One job per tool — update_order is better than update_anything. Narrow tools are easier for the model to use correctly.
- Idempotent where possible — repeat calls with the same arguments should be safe. Retries happen.
- Permission-scoped — read tools and write tools should be separate; high-stakes actions should require explicit approval.
- Robust error returns — when a tool fails, return a structured error the model can reason about and recover from, not a stack trace.
- Validation in the tool, not the prompt — never trust that the LLM will give valid arguments; validate inside the tool and return a clear error if not.
- Audit logging — every tool call should be logged with arguments, result and timing. This is your debugging surface when agents misbehave.
For a US engineering team in 2026, custom tools are where most of the AI product engineering happens. The LLM is roughly the same across products; what differentiates is the set of tools it can call into your specific system. A well-crafted set of 10–20 custom tools, exposed through MCP or the provider's tool-calling API, can transform an LLM from a chat box into a genuinely useful operator inside your business. Most of the operational care goes into making those tools safe, observable and correct — the LLM half is comparatively easy.
