The Last Mile to the Real World: Agent-Native Primitives as Reusable Building Blocks for Solving Real-World Tasks
HEELO raised completion on a 46-task real-world test from nine tasks per model to 39 and 35, the authors report.
The framework assembles task-specific graphs from reusable components, each pairing a resource with a resident language model that applies it in context. It checks results against evidence from the environment and revises implicated components after failures. Across more than 20 benchmarks, the authors report gains including Terminal-Bench 4.0 resolution rising from 37.3% to 55.5% with the same underlying model. On the 46-task test, they also report lower completion times and a 1.55-fold reduction in cost per completed task.
HF Daily Papers' note
The framework assembles task-specific graphs from reusable components, each pairing a resource with a resident language model that applies it in context. It checks results against evidence from the environment and revises implicated components after failures. Across more than 20 benchmarks, the authors report gains including Terminal-Bench 4.0 resolution rising from 37.3% to 55.5% with the same underlying model. On the 46-task test, they also report lower completion times and a 1.55-fold reduction in cost per completed task.
HF Daily Papers' note
score 5