LLM agents often fail at real-world tasks because they can access resources but don't know how to apply them correctly. For instance, they can retrieve a code snippet but not adapt it to the current problem. This gap is the "last mile."
The authors built reusable "agent-native primitives," each combining a resource with its own language model that knows how to use that resource. A framework called HEELO selects the needed primitives for each task, composes them into a sequence, and revises only the steps that fail.
On a benchmark, GPT-5.6 with its standard tool completed 37.3% of tasks, but adding HEELO with primitives raised success to 55.5%. This shows that embedding application knowledge with resources lets agents handle more real-world tasks efficiently.