Synthetic Sciences, a Y Combinator Winter 2026 (YC W26) company, has officially released OpenScience, an open-source research agent for scientific work. The system can search papers, process data, write code, and access scientific databases. It also adds a new feature called Autoresearch, which lets the agent keep running experiments on its own.
Autoresearch runs iterative experiments from a target
For tasks such as model training, a researcher only needs to give OpenScience a goal, such as lowering validation loss. The agent then runs a baseline, proposes modifications, and executes experiments round by round.
If a result improves, OpenScience keeps the change. If performance worsens, it rolls back the change and decides what to try next based on earlier results. Users can also set limits in advance, including the number of runs, total runtime, and stopping conditions. They can narrow the search space as well, for example by telling the system to change only the optimizer in the next steps.
Local and remote execution options
The release is aimed at automating the repetitive loop that researchers often manage manually: proposing changes, running experiments, comparing metrics, discarding failed attempts, and moving to the next round.
Experiments can run on a local machine or be sent to remote GPUs, SSH servers, and Slurm/PBS clusters. The code, results, and decision made in each round are recorded, making later inspection possible.
OpenScience also supports direct sign-in through ChatGPT/Codex subscriptions. Users can connect their own API keys, use local models, or choose Ace, the company’s official pay-as-you-go option.
Self-test results
In the company’s self-test, OpenScience completed 53 out of 70 Terminal-Bench Science tasks, for a score of 75.7%. For comparison, the public score for Codex + GPT-6 Astra was 68.1%.