Loading
An interface built for a human assumes a human decided to call it. An agent has to make that decision itself, hundreds of times, against a budget it cannot see.
The Model Context Protocol lets an AI assistant call your tools directly. Most enrichment products now ship an MCP server, and most of them are a thin wrapper over the same REST endpoints a developer would call — which is exactly where agents get into trouble.
An API designed for a human assumes a human decided to call it. An agent has to make that decision itself, hundreds of times, with a budget it cannot see and consequences it will not feel. Tool design is where that either works or quietly costs a fortune.
This is not a theoretical concern for us. We run a fine-tuned model with dozens of registered tools on our own hardware, and the single hardest part of that work was never adding capability — it was teaching the model to leave a tool alone.
Tool descriptions carry explicit negative guidance, not just capability. The model is told which questions this tool does not answer, so it stops reaching for it out of optimism.
Each tool declares what a call will consume, so an agent can be given a budget and reason about it rather than discovering the limit by hitting it.
Re-asking the same question returns the cached answer rather than re-billing it. Agents retry constantly; a tool that charges for every retry is a tool that punishes normal behaviour.
Not found, out of budget, provider down and malformed input are four different outcomes, and a model can only recover correctly if it can tell them apart.
Ask what a job would cost and which providers it would touch, without spending anything. Useful for an agent planning a batch, and for a human who wants to sleep at night.
To be clearNone of this exists yet. It is how we intend to build it, published now so the people who would use it can tell us where we are wrong before it is expensive to change.
Everything the agent surface can do is a plain request underneath. If you are wiring this into a pipeline rather than an assistant, that is the page you want.