Researchers say giving AI agents reusable 'primitives' helps them finish more real-world tasks

The problem: the last mile

As arxiv.org reported in a new paper, even the strongest AI models, each paired with its maker's own agent software, still miss a lot of real-world work. On Terminal-Bench 4.0, a test of tasks done in a command-line environment, the GPT-5.6 Sol model running in Codex resolves 37.3% of tasks. Claude Opus 5 running in Claude Code resolves 51.8%.

The authors argue these failures cluster at one specific point. The agent can often get to the right resource, such as a tool, a website or a file system. What goes wrong is the next step, actually applying that resource to the job. They name this gap the last mile to the real world.

A harness, in this context, is the scaffolding wrapped around a model: the instructions, tools and loops that let it act rather than just chat.

What an agent-native primitive is

The paper says existing approaches make a resource reachable but leave the know-how for using it to whoever is calling. So every team building a harness ends up re-creating that know-how from scratch.

The authors' answer is the agent-native primitive: a reusable unit that pairs a resource with a resident language model. That built-in model's job is to apply the resource sensibly in context. Primitives can also communicate with one another.

Think of the difference between handing someone a power tool and handing them the tool along with a skilled assistant who knows how to use it and can coordinate with other assistants. The competence travels with the resource instead of being rebuilt each time.

How HEELO assembles and improves them

The team also built a framework called HEELO, which stands for Harness Engineering around Evolving Agent-Native Primitives for the Last Mile to Real-World Tasks. It builds a custom harness for each task from a library of primitives.

The process runs in four steps:

• It picks the primitives a task needs and links them into a task-specific graph.
• It checks the outcome against evidence from the environment.
• When something fails, it revises the primitives implicated in the failure.
• It admits new or improved primitives to the library only after validating them.

The library is therefore meant to grow and get better over time. The authors built eight families of primitives, each aimed at one kind of task, and tested them on more than twenty benchmarks.

What the numbers show

The headline gain comes from Dev-Primitives. With the same underlying model as the Codex setup, they raised Terminal-Bench 4.0 resolution from 37.3% to 55.5%. That result passes the 51.8% the paper lists for Claude Opus 5 with Claude Code.

Other families showed gains on other tests:

• Harness Primitives, composed per task, lifted Terminal-Bench 2.0 Pass@1 from 60.0% for the starting harness to 72.2%. Pass@1 is the share of tasks solved on the first try.
• Tool Primitives raised the ToolBench pass rate from 55.8% to 75.1%.
• Navigation Primitives raised GAIA-Text-103 accuracy from 45.6% to 58.3%.

The authors also ran 46 real-world tasks on their own platform, heelo.dev. Working alone, a Claude Code setup using Claude Opus 5.5 and a Codex setup using GPT-6.1 Sol each finished 9 tasks. Connected to HEELO through an MCP server, a kind of plug that lets assistants call outside tools, they finished 39 and 35 respectively.

The paper adds that time per completed task fell by factors of 1.9 and 2.3, and cost per completed task was 1.55 times lower.

How much to trust it

This is a first-version preprint, so other researchers have not yet reviewed it. The most dramatic comparison, 9 tasks versus 39 and 35, comes from a platform run by the same authors who built the framework. That does not make it wrong, but it is not an independent test.

The standard benchmark results are easier to compare, since Terminal-Bench, ToolBench and GAIA are used by many groups. Still, the source does not say how the baselines were tuned or how much effort went into the primitives themselves. Those details would affect how far the gains would carry to your own work.

The idea is worth noting either way. It suggests that reusable, self-improving building blocks around a model may matter alongside raw model strength.