Large language models (LLMs) augmented with external tools have shown strong ability to solve complex real‑world tasks. Existing methods, however, suffer from two major issues: brittle multi‑step, multi‑turn reasoning due to mismatched tool output types and API schemas, and performance drop when the tool catalogue grows large. To tackle these problems we introduce Tool Primitives, which replace rigid API‑schema invocation with natural language. Each tool is wrapped by an LLM interface that resolves the schema and executes internally, enabling natural inter‑tool communication for nested and multi‑turn calls. Building on this, we create ToolFace, a centralized repository containing 25,519 functions. At inference time the LLM dynamically retrieves only the relevant tools, eliminating the need to enumerate raw API schemas in the prompt. To reliably orchestrate Tool Primitives and ToolFace in complex settings we propose HEART (Harness Engineering via Agent‑native Reusable Tool primitives), comprising a Planner, Router, and Verifier that jointly handle dynamic tool‑invocation planning, multi‑step execution, and feedback‑driven recovery. Experiments on five benchmarks show HEART outperforms SFT baselines by an average $10\%$, surpasses GPT‑5.4, Claude‑4.6‑Sonnet, and Gemini‑3.1‑Pro by $6\%$ on average, and cuts API cost by up to $85\%$. On 50 real‑world tasks HEART achieves $84\%$ task completion, which is $3.8\times$ the average of three leading commercial models (only $22\%$).
Review