Tool calling infrastructure for AI agents

AIThis description was written by a model from public signals. The facts behind it are listed under Evidence.

Problem

Open-source language models struggle with tool calling and function invocation compared to closed-source models, but the problem is often in the harness and validation layer rather than the model itself. Teams building AI agents need infrastructure to repair malformed tool inputs and schema validation errors.

Solution concept

A tool-input repair layer that sits between the language model and the validation schema. Instead of rejecting malformed inputs, it would automatically repair common schema mistakes, handle edge cases like markdown auto-links in file paths, and learn from recurring failure patterns. This would allow open models to match or exceed closed-source model performance on tool-calling tasks.

Who has it

Engineering teams building AI coding agents and autonomous systems using open-source language models like DeepSeek.

How they handle it today

Today, teams validate tool inputs against schemas (like Zod) and reject invalid calls, or preprocess inputs before validation. They encounter recurring schema mistakes and formatting issues that cause tool calling failures, requiring manual debugging and workarounds.

Why now

Open-source models like DeepSeek V4 Pro are becoming competitive with proprietary models on underlying capabilities, but tool-calling reliability remains a barrier. Teams want to switch to cheaper open models but need infrastructure to make them work reliably in production.

How it could charge

SaaS API for tool-calling repair with usage-based pricing (per request or per million tokens), or sold as a library with per-seat or per-deployment licensing for teams building internal AI agents.