RELAI optimized the harnesses around three open-weight agents, raising average pass rate from 13.3% to 38.7% while cutting average inference cost by 51%. That brought them to roughly frontier-level performance at ~3% of the inference cost, without changing model weights. The key insight: different models failed in different ways, so each needed a different harness—not a universally better prompt.

We optimized the agent harnesses around GPT-5.5, Claude Opus 5, and Gemini 3.5 Flash to improve quality and inference cost at the same time. Across 80 unseen tasks, cost per successful task fell by 22–44%, while pass rate improved for all three agents. Most of the savings came from carrying less unnecessary context, not simply making the agents do less.

RELAI was the only optimizer in our two-phase Terminal-Bench evaluation to both transfer to unseen tasks and keep improving after a second optimization round. Its lifelong average pass rate reached 76.4%, versus 66.0% for GEPA, 64.6% for Meta Harness, and 58.7% for the baseline.

June 10, 2026 — Today, we’re launching RELAI, a verifiable continual learning platform for AI agents, and announcing $6.9 million in total funding to scale our mission. The funding includes a newly secured $5.4 million pre-seed round led by .406 Ventures with participation from AITFund and other strategic investors, along with $1.5 million in prior investment support from Non sibi Ventures and TEDCO.

The frontier for AI agents is no longer whether they can use tools or pass a benchmark — it is whether they can learn continuously from experience. Four principles for a learning engine that improves without forgetting.