The Last Mile to the Real World: Agent-Native Primitives as Reusable Building Blocks for Solving Real-World Tasks

The Last Mile to the Real World. Agent-Native Primitives as Reusable Building Blocks for Solving Real-World Tasks. By Haibo Jin and thirteen others, from the University of Illinois Urbana-Champaign, the Georgia Institute of Technology, and one other institution. Published on the fifth of November.

How does an agent handle a real-world task from start to finish? Frontier models running in their own harnesses still fall short on real-world benchmarks. On Terminal-Bench 4.0, GPT-5.6 Sol with Codex resolves thirty-seven point three percent of the tasks, and Claude Opus 5 with Claude Code resolves fifty-one point eight percent.

The authors locate where these tasks fail: at the step from reaching a resource to applying it. They call that step the last mile to the real world. Existing approaches put the resource within reach, but they leave the competence to apply it with the caller. The consequence is that every harness has to rebuild that competence for itself.

The proposal is to move the competence into the resource. An agent-native primitive is a reusable unit that equips a resource with a resident LLM. That LLM applies the resource in context, and it communicates with other primitives. The resource no longer waits to be used correctly. It arrives with something that knows how to use it.

HEELO, short for Harness Engineering around Evolving Agent-Native Primitives for the Last Mile to Real-World Tasks, is the framework that puts primitives to work. It builds each task's harness from a library of them. For a given task, it composes the primitives that task needs into a task-specific graph. It verifies the result against evidence drawn from the environment. When something fails, it revises the primitives the failure implicates. And it admits new or improved primitives into the library only after validation, so the library evolves as tasks are solved.

HEELO is instantiated with eight primitive families, each built for one kind of task, and evaluated on more than twenty benchmarks.

Dev-Primitives raise Terminal-Bench 4.0 resolution from thirty-seven point three percent under Codex to fifty-five point five percent, on the same backbone. Harness Primitives, composed per task, raise Terminal-Bench 2.0 Pass@1 from sixty percent for the seed harness to seventy-two point two percent.

Tool Primitives raise the ToolBench pass rate from fifty-five point eight percent to seventy-five point one percent. Navigation Primitives raise accuracy on GAIA-Text-103 from forty-five point six percent to fifty-eight point three percent.

The authors also run forty-six real-world tasks on their heelo.dev platform. On their own, Claude Opus 5.5 with Claude Code and GPT-6.1 Sol with Codex each complete nine. Connected to HEELO through the authors' MCP server, Claude Opus 5.5 completes thirty-nine and GPT-6.1 Sol completes thirty-five. The time taken falls by one point nine times and by two point three times respectively, and cost per completed task is one point five five times lower.

The argument rests on a division of labour. Existing approaches already bring the resource within reach, and applying it is where the tasks were breaking. Giving every resource its own resident model, and letting a harness assemble and revise those units per task, means the competence is built once and reused, instead of being rebuilt inside each harness that needs it.