The last mile to the real world: agent-native primitives as reusable building blocks for solving real-world tasks

The paper proposes agent-native primitives and the HEELO scaffolding engineering framework, locating an agent's failure on real tasks at the step from acquiring resources to applying resources, letting resources carry a resident LLM that applies them in context and communicates with other primitives, composing primitive graphs per task, and verifying and revising with environmental evidence. The paper instantiates eight primitive families and evaluates on more than twenty benchmarks: Dev-Primitives raises Terminal-Bench 4.0 solve rate from 37.3% to 55.5% on the same GPT-5.6 Sol with Codex backbone, Harness Primitives raises Terminal-Bench 2.0 Pass@1 from 60.0% to 72.2%, Tool Primitives raises ToolBench pass rate from 55.8% to 75.1%, and Navigation Primitives raises GAIA-Text-103 accuracy from 45.6% to 58.3%.