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

LLM agents have become increasingly capable at reaching tools, repositories, documents, and other agents. Yet reaching a resource is not the same as using it correctly. A tool may require arguments grounded in an earlier result, a code component may need adaptation to its target repository, and a failed action may need to be traced back to the specific work that caused it.

The authors call this gap the last mile to the real world. They argue that current agent harnesses often leave the competence for applying a resource with one central caller. As a result, every new task or harness must reconstruct the same practical knowledge: how to bind inputs, satisfy dependencies, validate outcomes, and recover from failure.

From passive resources to agent-native primitives

The paper’s central proposal is an agent-native primitive: a reusable resource paired with a resident language model. The resource might be an API schema and function, a code component, a proof method, a source model, a multi-agent pattern, or a repository file. Its resident model applies that resource to a natural-language request in context and can send requirements or useful information to other primitives.

This changes the division of labor. Instead of exposing every implementation detail to a task-level model, the primitive owns the knowledge needed to use its resource. A tool primitive resolves natural-language requests into typed API arguments. A code primitive adapts its component to a target repository and runs its carried tests. A math primitive describes when a proof method applies and what obligations remain. In domains with exact procedures, implemented code can replace the resident model.

HEELO builds a harness around demand

The authors introduce HEELO, short for Harness Engineering around Evolving Agent-Native Primitives for the Last Mile to Real-World Tasks. It builds a task-specific harness from a growing library through three mechanisms.

  1. Selection and composition: HEELO finds candidate primitives, assesses their relevance, and connects the few needed for a particular task into a graph. Results and messages flow along explicit dependencies.
  2. Evidence-based execution: The graph runs in the task’s environment. Tests, compilers, runtime traces, source evidence, or other acceptance checks determine whether the integrated result passes.
  3. Targeted revision and publishing: When a failure occurs, HEELO attempts to identify the implicated primitives and revises only that work. New or improved primitives enter the reusable library only after separate validation.

This design leads to a useful theoretical claim: if activation is driven by unresolved task needs, the number of active primitives can be bounded by task demand rather than by the total library size. The authors also analyze when task-specific composition can outperform a fixed harness, although that result depends on simplifying assumptions about additive gains, penalties, and valid execution.

Results across eight primitive families

HEELO is instantiated through eight families: Tool, Code, Math, Knowledge, Agent, Dev, Navigation, and Harness Primitives. Across more than twenty benchmarks, the authors report improvements over the relevant reference systems.

  • Dev-Primitives raise Terminal-Bench 4.0 resolution from 37.3% under Codex to 55.5% in the reported maximum-effort comparison.
  • Harness Primitives raise Pass@1 from 73.0% to 80.5% on 100 sampled SWE-bench Verified tasks, and from 60.0% to 72.2% on 45 Terminal-Bench 2.0 tasks.
  • Tool Primitives raise ToolBench pass rate from 55.8% to 75.1% on a matched Qwen3-8B setting.
  • Code Primitives raise repository delivery from 0.3180 to 0.5134 on 522 repository-reconstruction tasks.
  • Agent Primitives raise mean accuracy from 58.8% for a single agent to 75.3% across eight math, code, and question-answering benchmarks in one reported Qwen3-8B setting.
  • Navigation Primitives raise GAIA-Text-103 accuracy from 45.6% to 58.3%.

The ablations are especially important. The reported gains generally disappear when the resource is merely retrieved and shown to a central caller. ToolBench performance falls sharply when Tool Primitives are removed, while repository delivery benefits primarily from adapting retrieved code rather than simply accessing it. In software engineering, removing diagnosis feedback produces one of the largest declines, supporting the importance of routing failures back to the components that own them.

A live platform demonstration

Beyond frozen benchmarks, the authors built heelo.dev, which exposes actions such as publishing pages, sending outreach, reading invoices, and navigating services through agent-native primitives. On 46 real-world tasks, Claude Code and Codex each complete 9 tasks alone. Connected through HEELO’s MCP server, they complete 39 and 35 respectively. The reported overall task scores rise from 28% to 90% for Claude Code and from 25% to 83% for Codex. Time and cost per completed task also fall in the authors’ accounting, because the connected agents finish substantially more tasks.

These results demonstrate the value of giving an agent controlled access to capabilities it otherwise cannot exercise. They do not show that HEELO solves arbitrary real-world automation: the platform uses defined task criteria, test accounts, allowlisted recipients, and a particular collection of tools.

Limitations and practical implications

The evaluation combines eight companion studies with different models, datasets, metrics, and baselines. Some comparisons are matched-backbone ablations, while others compare against systems with their own configurations, so headline gains should not be treated as one uniform measurement. The real-world platform evaluation is also limited to 46 tasks and depends on the availability and correctness of its integrations.

There are practical costs. Agent-native primitives add model calls, library construction, validation, provenance tracking, and another layer of operational complexity. Selection can miss the right primitive, and verification can produce false passes or false failures. The paper’s own grounding audit also reports that some learned source-model units are only partially supported or unsupported, reinforcing that a source model should complement—not replace—the authoritative source.

For teams building agents, the most actionable lesson is architectural: package reusable capabilities with the logic required to apply them, give each capability explicit interfaces and validation, route outputs to their consumers, and make failures diagnosable. HEELO suggests that the next step beyond adding more tools may be engineering better reusable units around those tools. That approach is demonstrated across the paper’s tested settings; whether it becomes a broadly reliable solution for open-ended real-world work remains an empirical question.


Sources

AI-assisted research summary. Check the original paper for complete methods, results, and limitations.

Avatar photo

Tech News

Jeff is a passionate blog writer who shares clear, practical insights on technology, digital trends and AI industries. With a focus on simplicity and real-world experience, his writing helps readers understand complex topics in an accessible way. Through his blog, Jeff aims to inform, educate, and inspire curiosity, always valuing clarity, reliability, and continuous learning.

↑ ↓ navigate Enter open Esc close