The paper addresses the "last mile" problem where frontier models fail to apply accessible resources to real-world tasks, evidenced by GPT-5.6 Sol with Codex resolving only 37.3% and Claude Opus 5 with Claude Code 51.8% on Terminal-Bench 4.0. The authors introduce agent-native primitives—reusable units pairing a resource with a resident LLM that applies it contextually and communicates with other primitives—and HEELO (Harness Engineering around Evolving Agent-Native Primitives for the Last Mile to Real-WOrld Tasks), a framework that composes task-specific harnesses from a primitive library. HEELO constructs a task-specific graph of primitives, verifies execution against environmental evidence, revises implicated primitives upon failure, and admits new primitives post-validation. Instantiated with eight primitive families across over twenty benchmarks, Dev-Primitives lift Terminal-Bench 4.0 resolution from 37.3% to 55.5% on the same backbone; Harness Primitives raise Terminal-Bench 2.0 Pass@1 from 60.0% to 72.2%; Tool Primitives improve ToolBench pass rate from 55.8% to 75.1%; and Navigation Primitives increase GAIA-Text-103 accuracy from 45.6% to 58.3%. On the authors' 46-task real-world platform, Claude Opus 5.5 and GPT-6.1 Sol complete 9 tasks independently but 39 and 35 respectively when connected to HEELO via an MCP server, achieving 1.9× and 2.3× speedups and 1.55× lower cost per completed task.