Netflix has released oci-agent, a new open-source tool designed to streamline complex observational causal inference while keeping human analysts involved in key decisions. The project aims to automate repetitive and error-prone parts of causal analysis without replacing the judgment required to frame research questions and interpret results.
Netflix Releases oci-agent as Open-Source Causal Analysis Tool
Netflix officially open-sourced oci-agent with its initial v0.1.0 public release through the Netflix-Skunkworks repository on GitHub.
The system is designed for Observational Causal Inference (OCI), a field focused on estimating causal relationships using observational data rather than controlled experiments. Such analysis can be particularly valuable when conducting a traditional randomized A/B test is impractical or impossible.
For Canadian researchers, data scientists and technology teams working in areas ranging from digital services to public policy, causal inference methods can help distinguish correlation from evidence of a potential causal effect.
Netflix says oci-agent is intended to automate labour-intensive elements of the process, including sensitivity analysis and the management of multiple analytical iterations. Higher-level responsibilities, such as defining the research question and evaluating the results, remain with human analysts.
Dual-Agent Actor-Critic Architecture Drives the Workflow
A central feature of oci-agent is its dual-agent actor-critic architecture. The approach divides responsibilities between two agents that perform different roles during an analysis.
Actor Agent Executes the Causal Analysis Plan
The actor agent carries out the analytical workflow. Human analysts begin by providing an analysis plan along with a templated Jupyter notebook.
Using those inputs, the actor produces a specification, fills in the required notebook parameters and executes the analysis code.
The OCI workflow is framed around target trial emulation. In practical terms, causal estimation is approached as an effort to identify the most appropriate hypothetical A/B test for the research question and then emulate that experiment using observational data.
Critic Agent Reviews Results and Identifies Weaknesses
The critic agent evaluates the actor’s work, looking for potential gaps, methodological problems or weaknesses that could affect the reliability of the findings.
This review mechanism is especially important in observational causal inference because analysts often lack a known ground truth against which an estimate can be directly evaluated.
Rather than relying solely on a final numerical result, oci-agent combines human oversight with a transparent and auditable analytical process.
Inspectable Artifacts Keep Humans in the Loop
Netflix’s approach emphasizes reproducibility and reviewability. The agents generate artifacts that analysts can inspect and execute again, including analysis plans, specifications, plots and updated Jupyter notebooks.
This design allows researchers to examine how an estimate was produced instead of treating an AI-generated result as a black box.
Human analysts therefore remain responsible for important decisions, particularly when assessing assumptions, interpreting findings and determining whether the evidence supports a meaningful causal conclusion.
Benchmark Testing Shows Competitive Performance
Netflix tested the workflow using data from the Atlantic Causal Inference Conference (ACIC), a widely used setting for evaluating causal inference methods.
According to Netflix, oci-agent achieved competitive performance compared with existing systems in those benchmark tests.
The company also examined the workflow in a case study involving the estimated effect of introducing new entertainment types, such as games, on customer retention.
Critic Agent Flags Early-Adopter Bias
In the Netflix case study, a baseline Claude model using simple linear regression produced an overestimate of the retention effect.
The oci-agent workflow generated a substantially lower estimate — just 25 per cent of the baseline figure — after the critic agent identified early-adopter bias and a failed placebo test.
The example illustrates the purpose of the actor-critic structure: rather than simply automating calculations, the system is designed to challenge analytical assumptions and surface problems that could materially change a causal estimate.
Open-Source Release Expands Access to Causal Inference Workflow
By making oci-agent publicly available, Netflix is opening its experimental causal inference workflow to researchers, developers and data science teams interested in combining AI agents with reproducible statistical analysis.
The initial v0.1.0 release remains an early version, but its emphasis on transparent artifacts, automated analytical tasks and human oversight offers a model for applying AI to complex causal research. For organizations working with observational data, the project demonstrates how agent-based systems can support analysts while keeping critical methodological decisions subject to human review.
William Faulkner was a Nobel Prize-winning American novelist and short story writer renowned for his innovative literary techniques and profound exploration of the American South. His works remain among the most influential contributions to twentieth-century literature.