The hard part of real-world agent work is often not finding a tool, code snippet, or document—it is applying it correctly to the task at hand and carrying its result through the next steps. This paper’s central insight is to package that application competence with the resource, then let the task assemble only the components it needs.
Key Findings
- Agent-native primitives pair a resource with a resident LLM. This lets each unit interpret requests in context and communicate its requirements and results to other units, rather than leaving every caller to rebuild that expertise.
- HEELO composes task-specific graphs and checks outcomes against evidence from the environment. When something fails, it traces the failure to implicated primitives, avoiding unnecessary repetition of successful work.
- The framework spans eight primitive families, including tools, code, math, knowledge, multi-agent systems, software engineering, navigation, and harnesses. Across more than twenty benchmarks, the paper reports gains in several settings; for example, Dev-Primitives reach 55.5% on Terminal-Bench 4.0 versus 37.3% for Codex on the same backbone.
- The real-world evaluation suggests practical value beyond benchmark tasks. On 46 tasks, Claude Code completes 39 with HEELO versus 9 alone, while Codex completes 35 versus 9 alone. These results are specific to the tested platform and models, rather than a guarantee for arbitrary workflows.
How the Approach Works
The library evolves between tasks: candidate primitives are selected and composed for a particular request, their outputs are routed to downstream consumers, and evidence from execution guides targeted revision. New or improved units enter the library only after validation, such as tests, compilation, or development runs. This makes reusable competence—not a fixed, monolithic harness—the unit that accumulates over time.
Who It’s For
A good fit for researchers and engineers building agents that must coordinate tools, code, and other resources across multi-step tasks, especially when failures need to be diagnosed and corrected selectively. Look elsewhere if you need a drop-in agent for a single narrow task, or if your priority is a simple, fully specified runtime: the approach depends on designing, validating, and maintaining a growing primitive library, and its reported gains come from particular benchmarks and experimental setups.