Cobalt is a cloud-based coding agent service released by Amaretto Software Labs Ltd on July 21, 2026. When developers delegate tasks, Cobalt provides isolated, persistent cloud computers for each agent, where it installs project dependencies, runs databases, modifies code, and executes tests before delivering results as GitHub commits and Pull Requests (PRs). While typical code generators focus on presenting suggestions as text or patches, Cobalt aims to execute and validate code in a real application environment, then return the changes as reviewable units for human inspection. This approach is akin to a cooking assistant that prepares dishes in a separate kitchen, checks their status, and serves the finished product, rather than merely providing recipes. Results from existing AI coding tools often require developers to re-verify them due to differences in execution environments, missing dependencies, database connection issues, or failed tests. This is particularly true for long-running tasks, where reconstructing the environment after a session ends can become repetitive. Cobalt’s differentiator lies in assigning isolated persistent computers per agent to maintain the working environment and progress state, extending beyond the suggestion phase to include execution and testing. Since the final output takes the form of GitHub commits and PRs, developers can review changed files, test results, and code diffs within their existing review workflows rather than manually transferring conversation content. However, technical details regarding the isolation method, supported languages and frameworks, security and data retention policies, pricing plans, and self-hosting support could not be confirmed based on the available discovery information. Biotech software researchers can leverage Cobalt to improve reproducibility of analysis pipelines or handle repetitive maintenance tasks. For example, one could configure a workflow that installs dependencies in a Python-based genomics analysis repository, runs a test database, modifies a variant filtering module, and receives only the results that pass unit tests as a PR. In a microscopy image processing project, researchers can request input format validation, batch processing code modifications, and test execution as a single task, reviewing the changed behavior at the commit level. Such use cases represent a general coding agent scenario where the researcher delegates development and verification tasks to Cobalt within their own managed code repository, rather than implying that Cobalt provides specific bioinformatics functionalities natively. When applying Cobalt to repositories containing research data, separate review is necessary. Since code and data may be processed on external cloud computers, it is essential to verify the service terms of use, privacy policy, data storage location, retention period, whether data is used for training, and deletion procedures before connecting patient-derived information, unpublished sequences, or regulated data. Furthermore, even when PRs are generated, scientific validity and statistical accuracy are not automatically guaranteed; therefore, an operational approach that utilizes synthetic test data and CI (Continuous Integration) checks, with final review by domain researchers, is appropriate.

Cobalt is a cloud-based coding agent service released by Amaretto Software Labs Ltd on July 21, 2026. When developers submit tasks, Cobalt provides isolated persistent cloud computers for each agent, where it installs project dependencies, runs databases, modifies and tests code, and then delivers the results as GitHub commits and Pull Requests (PRs). While typical code generators focus on presenting suggestions as text or patches, Cobalt executes and validates code in a real application environment before making it available for human review.

🔬 Genomic Analysis Code Regression Verification: Request modifications to the variant filtering module and testing in a Python genomic analysis repository, setting the approval condition as 100% pass rate for three synthetic VCF fixtures and existing unit tests. After the researcher reviews the generated GitHub PR, reflect the changes in a reproducible analysis pipeline.