Jev in practice: typed decisions, scoped authority

Safe agent delegation with Jev, LangGraph, and Tenuo represents a structured approach to upgrading software dependencies, ensuring that each step within the process is carefully controlled and auditable. Jev, a System One Model for decisions, transforms repository evidence into typed probabilistic choices. LangGraph carries these decisions through a durable workflow, while Tenuo assigns task-scoped authority to each action.

The key principle is to separate judgment, control flow, and authority. The design process began with the question of what a practical workflow would look like when each component addresses distinct aspects of the system. Dependency upgrades provided a practical context, highlighting the challenges of seemingly mechanical processes that can conceal complex compatibility issues.

Upgrading dependencies involves changing versions, installing packages, running tests, and initiating pull requests, but the release may alter APIs, module formats, or behaviors that existing tests do not cover. A robust upgrade agent must interpret release evidence, relate it to unfamiliar repositories, decide required actions, implement changes, and verify the results.

It also requires access to source files, package installation, test commands, Git, and sometimes remote repositories. Jev excels at making useful decisions from complex contexts, providing developers with typed options and confidence scores. The workflow then uses these decisions within a structured graph, with each node receiving explicit state and returning updates.

Execution begins when the selected action is mapped to a trusted worker, establishing clear authority boundaries for each task. LangGraph tracks repository facts, findings, decisions, attempts, and evidence throughout the run, ensuring accountability and traceability. Tenuo provides task-level authority, narrowing the scope of actions a worker can perform, ensuring safe and controlled execution.

Employing this architecture in CI environments like GitHub Actions allows for granular control, limiting the scope of modifications and ensuring that only authorized actions are taken. This structured approach to agent delegation enhances safety, control, and accountability in software upgrades.