OpenAI Report: Coding Agents Help Scientists Build Faster Research Software
OpenAI has published a field report showing how AI coding agents helped researchers improve, rebuild, and accelerate scientific software used in genomics, immunology, statistics, and RNA sequencing.
The report examined eight scientific computing projects. Five used OpenAI Codex, while three combined Codex with Anthropic’s Claude Code.
However, the findings should be viewed with proper context. OpenAI is reporting on the use of its own technology, and the case studies were written by researchers and developers who participated in the projects. The report is not an independent scientific evaluation.
Still, the results highlight a growing role for AI agents in addressing one of the biggest challenges in scientific research: maintaining software after the original research project has ended.
Scientific software often suffers from technical debt
Many research tools are created by small academic teams to support a specific study or scientific paper. These teams may not have dedicated software engineers, long-term funding, or enough time to update aging code.
As a result, valuable scientific software can become difficult to install, slow to run, incompatible with modern systems, or completely abandoned.
OpenAI’s report suggests coding agents can help researchers modernize these tools by automating repetitive engineering work, proposing optimizations, migrating programs to newer systems, and rewriting software in faster programming languages.
The eight projects focused on three main areas:
- Updating packaging and build systems
- Improving the performance of existing software
- Rebuilding tools using new programming languages or computing backends
AI agents improve genomics and sequencing tools
One project involved cyvcf2, a Python library used to read genomic variant files. Its outdated packaging and build system was replaced with a modern, unified process.
Contributor Brent Pedersen said AI agents can help developers move quickly, but meaningful scientific progress still requires human expertise, judgment, and care.
Another project focused on HI.SIM, a simulator for DNA-sequencing reads.
Two mostly autonomous optimization attempts using GPT-5.2 and GPT-5.6 reportedly reduced the software’s runtime by 31 percent across a representative test set without changing its output.
Contributor Andrew Ho said he was not a specialist in genomics or the C programming language. From his perspective, the coding agent helped solve performance and packaging issues that he could identify but did not personally know how to repair.
The genome assembly tool Hifiasm also received performance improvements. The agent reportedly reduced runtime by 25 percent on its main optimization target and by approximately 15 percent when tested on separate human sequencing data.
Contributor Suyash Shringarpure said the agent created its own benchmarking environment and independently suggested possible improvements.
However, he emphasized that human involvement remained essential. Researchers still needed to provide profiling data, review results, and guide the agent away from repeated mistakes.
AI migrates scientific software to modern platforms
Coding agents were also used to modernize MHCflurry, a machine-learning tool that predicts which protein fragments may be presented to immune system T cells.
The project migrated the software from TensorFlow and Keras to PyTorch while maintaining compatibility with previously released model weights.
Contributors Alex Rubinsteyn, Sergey Feldman, and Timothy O’Donnell described this kind of work as important but often overlooked. Updating software backends may not attract the same attention as new scientific discoveries, but it can prevent valuable open-source tools from becoming outdated or unusable.
Another project, bayesm-rs, recreated statistical models from the R-based bayesm package using the Rust programming language.
The Rust version reportedly produced estimates within an established tolerance of the original software. It also ran 2.3 to 2.7 times faster using a single processor thread and 4.4 to 9.5 times faster across eight threads.
Contributors Andrew Bai and Andrew Ho said the agents performed well when they had a clear reference implementation to compare against.
However, tasks involving statistical assumptions or scientific judgment still required direct human review.
Rust rebuilds deliver major speed improvements
Several projects used AI agents to rebuild scientific applications in Rust, a programming language known for speed, reliability, and memory safety.
One project, rustar-aligner, recreated STAR, a widely used RNA-sequence alignment tool that had lost active maintenance.
Contributor James M. Ferguson said rewriting a 20,000-line program manually would not normally be considered a practical use of a researcher’s time. With coding agents, however, the task became possible through several weeks of guided development.
Verification remained a major challenge.
According to Ferguson, an AI system could claim that an output graph looked correct, but researchers still had to manually inspect more than 900 plots before releasing the software.
Another project, RustQC, combined 15 separate RNA-sequencing quality-control tools into one application.
Contributor Phil Ewels said the new program reduced runtime by approximately 60 times and lowered disk input and output operations by 25 times.
Related rebuilds also produced significant performance gains. FastQC-Rust reportedly ran seven times faster, while a rebuilt version of Trim Galore ran three times faster while maintaining the original tools’ behavior.
Ewels warned that easier software rebuilding could also create new problems.
When different laboratories release multiple versions of the same scientific tool, those versions may behave differently. This can make research results difficult to compare across institutions or over time.
The engineering may become easier, but deciding who will maintain, govern, and support the software remains a serious challenge.
GPU-based redesign accelerates mutation simulation
The report also featured HelixForge, a GPU-native rebuild of BAMSurgeon, a tool used to simulate genetic mutations.
According to contributors Mamad Ahangari, Varun Goyal, and Hassan Masoudi, HelixForge reduced runtime by around 60 times in a benchmark using real human genomic data.
The team also reported that the rebuilt system generated mutation frequencies closer to the requested targets and fixed several bugs that caused unwanted artifacts in the original software.
Scientific verification remains the biggest challenge
Across all eight projects, one pattern remained consistent: AI agents were effective at implementing clearly defined programming tasks, but they could not reliably determine whether their own output was scientifically valid.
Some contributors reported that coding agents expressed confidence even when their work contained obvious errors.
This shifted the most important responsibility back to humans.
Researchers had to design strong acceptance tests, compare results with established tools, verify exact outputs, analyze numerical differences, and test the software using simulations where the correct answer was already known.
In many projects, AI agents produced the first version of the code quickly. Most of the remaining development time was spent reviewing edge cases, investigating small numerical differences, and checking whether faster software still produced trustworthy scientific results.
Faster development creates opportunities and risks
Lower software engineering costs could allow small research teams to rebuild scientific tools that previously required large grants or specialized engineering staff.
At the same time, easier development could result in several teams creating competing and incompatible versions of the same software.
Some projects in the report were integrated into their original open-source communities. Improvements to cyvcf2 and MHCflurry were contributed back to the existing projects.
Rustar-aligner, meanwhile, was transferred to new community stewardship because the original software it replaced was no longer actively maintained.
The report’s strongest message is not simply that scientists should use AI coding agents.
Instead, research teams must decide who will own, maintain, validate, and govern an AI-assisted software project before releasing it to the scientific community.
AI may now generate code faster than ever, but human experts remain responsible for determining whether that code is accurate, reproducible, safe, and scientifically meaningful.
For researchers in the Philippines, coding agents could help universities, laboratories, and small development teams modernize aging research tools with fewer engineering resources. However, proper testing, data governance, scientific validation, and long-term software stewardship must remain central to every AI-assisted project.
