New framework improves how AI systems use external tools with natural language interfaces

A team of researchers has developed a new approach to help large language models (LLMs) more reliably work with external tools and services. The work addresses longstanding problems that arise when AI systems need to chain multiple tool calls together or access large collections of available functions.

The core problem#

Current systems that augment LLMs with external tools often struggle when tasks require several steps. The main issues stem from incompatible tool output formats and API schemas — when one tool's output doesn't fit cleanly into the next tool's expected input format, the reasoning chain breaks down. Additionally, performance tends to degrade significantly when an LLM must choose from large catalogs of available tools, since listing all options in the context window becomes unwieldy.

A new design: Tool Primitives#

Rather than requiring LLMs to navigate rigid API schemas, the researchers introduced Tool Primitives — a wrapper layer that presents tools through a natural language interface. Each tool is encapsulated with its own built-in LLM interface that handles schema resolution and execution internally. This abstraction enables tools to communicate with each other seamlessly, supporting nested and multi-turn interactions without constant reformatting.

Paired with this abstraction is ToolFace, a centralized repository holding 25,519 functions. Instead of listing every available function upfront, the system dynamically retrieves only the relevant tools during inference, keeping the context window lean and focused.

HEART framework#

To reliably orchestrate Tool Primitives and ToolFace at scale, the team designed HEART (Harness Engineering via Agent-native Reusable Tool Primitives). HEART comprises three key components:

• Planner — creates dynamic invocation strategies for multi-step tasks
• Router — directs tool calls efficiently through the system
• Verifier — monitors execution and triggers recovery actions when needed

Results#

Experiments across five benchmarks showed HEART outperforming supervised fine-tuning models by roughly and surpassing GPT-5.4, Claude-4.6-Sonnet, and Gemini-3.1-Pro by . On 50 real-world tasks, the system achieved — nearly than the average of three leading commercial models tested, which reached only 22%. The approach also reduced API costs by up to compared to baseline approaches.