A team of researchers has built a system that upgrades the scaffolding around an AI agent, not just the AI itself.
Described in a paper posted to arXiv in August 2026, HELIX is a framework for improving the "harness," the code that manages an agent's context, tool access, and decision-making, alongside the underlying model. The approach: build harnesses for a fixed model, use verified results from many parallel attempts to generate training data, then rebuild the harness again as the model improves. In one evaluation round on code-repair tasks, a portfolio of 65 candidate harnesses found a single configuration that improved task coverage by 4.0% over a baseline called Pi. Running the full portfolio together, rather than just the winning candidate, surfaced up to 58.0% more verified coverage by combining differing approaches, and a 200-slot slice of that output yielded 438 usable training records.
Most agent research treats the model as the bottleneck and the runtime around it as plumbing. HELIX argues the plumbing matters just as much: what tools an agent can call and when it decides to stop shapes both its current performance and what data becomes available to train the next model. That's a useful reframe for anyone building agent products, since it suggests gains can come from re-engineering the runtime as often as retraining the model.
This is one evaluation round on one benchmark, code repair scored with the SWE-bench evaluator, so the percentage gains are early signal rather than a settled result. The code is open-source on GitHub for anyone who wants to check the claims themselves.