From Tool Access to Tool Use: Agent-Native Primitives Target the Last Mile

Introduction

For an AI agent, having access to a terminal, browser, code repository, or external tool does not guarantee that it can complete a task. The difficult step often comes after the resource has been reached: understanding the context, choosing the right operation, and using the result correctly. Existing harnesses tend to rebuild this competence through prompts, scripts, and one-off orchestration, which makes systems difficult to reuse and improve.

A study featured by Hugging Face Daily Papers proposes Agent-Native Primitives as a way to address this gap. The idea is to turn the ability to apply a resource into a reusable component rather than leaving it entirely to the calling agent.

Key ideas

• Resources become capability-bearing units: An agent-native primitive combines a resource—such as a tool, codebase, or knowledge source—with a resident LLM-powered expert. That expert interprets the task context, applies the resource, and communicates with other primitives.
• Task-specific harnesses are composed dynamically: HEELO, short for “Harness Engineering around Evolving Agent-Native Primitives for the Last Mile to Real-World Tasks,” selects and connects the primitives required by a particular task.
• Verification drives improvement: The framework checks outputs against evidence from the environment. When a failure occurs, it identifies the primitive most likely involved, revises it, and admits a new or improved version only after validation.
• Multiple primitive families are evaluated: The authors instantiate eight families covering different task types, including development, tool use, and navigation, and test them across more than 20 benchmarks.

Results and implications

According to the paper’s reported results, Dev-Primitives raise Terminal-Bench 4.0 resolution from 37.3% with Codex to 55.5% on the same backbone. Task-composed Harness Primitives improve Terminal-Bench 2.0 Pass@1 from 60.0% to 72.2%. Tool Primitives increase the ToolBench pass rate from 55.8% to 75.1%, while Navigation Primitives raise GAIA-Text-103 accuracy from 45.6% to 58.3%.

The study also reports results on 46 real-world tasks hosted by heelo.dev. Claude Code and Codex each completed 9 tasks when operating on their own; after connecting to HEELO through its MCP server, the corresponding figures rose to 39 and 35. The abstract also reports lower time and cost per completed task. These numbers should be read as results from the authors’ evaluation setup, not as a universal guarantee across all agent deployments.

The broader significance is a shift in what gets reused in an agent stack. Conventional systems reuse APIs, tools, and prompts. Agent-Native Primitives aim to reuse the competence required to apply those resources in context. If the approach holds across independent tasks and models, agent engineering could move from continuously adding tools toward maintaining a verifiable and evolving library of specialized experts. Open questions remain around coordination overhead, the reliability of environmental evidence, and transfer across domains.

Source: Hugging Face Daily Papers