Netflix’s oci-agent is a dual-agent actor-critic workflow designed to automate complex causal data analysis while maintaining human-in-the-loop oversight.
Netflix officially open-sourced oci-agent on 28 May 2026 (v0.1.0 initial public release) on GitHub under the Netflix-Skunkworks repository for Observational Causal Inference (OCI). The tool is designed to automate repetitive or error-prone tasks in causal analysis, such as sensitivity analysis and tracking multiple iterations, leaving higher-level tasks like question framing and result evaluation to human analysts.
The oci-agent source code is publicly available on GitHub. The workflow frames OCI analysis as target trial emulation, treating causal estimation as finding the optimal A/B test for a given question. The system utilises a dual-agent actor-critic architecture, where an actor agent executes the analysis plan and a critic agent reviews the results and surfaces potential gaps or flaws.
Human analysts initialise the process by providing an analysis plan and a templated Jupyter notebook. The actor agent uses the plan to produce a specification, populate notebook parameters, and execute the analysis code.
To address the evaluation challenge in causal inference (where ground truth is absent), the workflow combines transparent process audits with human oversight. The agents generate inspectable and re-executable artifacts, including plans, specifications, plots, and updated Jupyter notebooks, rather than simply providing final output values.
In benchmark testing on the Atlantic Causal Inference Conference (ACIC) dataset, Netflix found the workflow achieved competitive performance compared to existing systems.
In a Netflix case study estimating the retention impact of new entertainment types (e.g. games), a baseline Claude model returned an overestimate via simple linear regression, whereas the oci-agent workflow produced an estimate that was just 25% of the baseline after the critic agent flagged early-adopter bias and a failed placebo test.
