OpenScience scores 75.7% in self-test, outperforming Codex: Research agents now run experiments autonomously in a loop.

Beating AI News (from Dongcha) reports that Synthetic Sciences, a Y Combinator Winter 2026 (YC W26) startup, has officially launched its open-source research agent OpenScience. The tool can search academic papers, process data, write code, and access research databases; its newly added Autoresearch feature allows AI to run experiments continuously on its own. For instance, when training a model, researchers only need to instruct it to “reduce the validation set loss”. OpenScience will first run a baseline, then propose modifications, and execute experiments iteratively. It retains improved results, rolls back those that worsen, and determines next steps based on prior outcomes. Users can also pre-set limits on run count, total duration, and stop conditions, or specify constraints like “only modify optimizers moving forward”. OpenScience automates the experimental iteration workflow that previously required researchers to monitor closely: proposing plans, running experiments, comparing metrics, discarding failed approaches, and proceeding to the next round. Experiments can run locally, or be deployed to remote GPUs, SSH servers, and Slurm/PBS clusters. Every round’s code, results, and judgments are logged for later review. The agent supports direct login to ChatGPT/Codex subscriptions, as well as integration with users’ own API keys, local models, or the official pay-as-you-go Ace. In official self-tests, OpenScience completed 53 out of 70 Terminal-Bench Science tasks, scoring 75.7%. By comparison, Codex + GPT-6 Astra posted a public score of 68.1%.