AgentConnect has introduced support for Jev, TypeSafe’s judgment model, transforming it into a general-purpose decision engine for AI agents. This integration allows developers to define reusable questions—termed 'Decisions'—that evaluate context to determine routing, categorization, or action triggers. Instead of relying on brittle if-else chains for every agent interaction, teams can now configure agents to answer specific criteria, such as whether a support message requires a reply or if a code issue needs a security review.

The Power of Reusable Decisions

The core feature, 'Decisions,' can return binary Yes/No answers, select from defined categories, or score inputs on a custom scale. For instance, a 'Needs reply' Decision can filter out greetings and acknowledgments in a support channel, ensuring agents only engage with substantive requests. These Decisions are reusable across multiple contexts; the same logic used to categorize a Slack message as 'Technical' or 'Billing' can be applied to route incoming issues in GitHub or GitLab to the appropriate reviewer.

Beyond Routing: Model Selection and Self-Querying

The utility of Jev extends beyond simple triage. AgentConnect allows teams to use Decisions for runtime and model selection, mapping task complexity scores to preferred LLMs. This enables cross-model review workflows, such as having Codex review a PR authored by Claude, and vice versa. Furthermore, agents can invoke Decisions themselves during execution via the Tools & Skills interface, allowing them to self-classify tickets or assess context before taking action, all while respecting existing permission boundaries.

Developer Workflow and Testing

Creating these Decisions is designed to be iterative and testable. Developers can start in a Playground conversation, describing the judgment they want the model to make. The agent drafts the Decision, which must be explicitly approved before application. Once saved, the Decisions page allows for 'Try with an example' testing, where users can inspect the model’s answer and probability distributions. This transparency is critical for debugging why an agent routed a specific query or declined to respond.

Key Takeaways

  • AgentConnect now supports TypeSafe’s Jev model as a central decision engine for agent orchestration.
  • 'Decisions' are reusable, testable logic blocks that can categorize, score, or filter agent inputs.
  • The integration supports complex workflows like cross-model code reviews and dynamic runtime selection.
  • Developers can test Decisions with sample conversations and inspect probability outputs before deployment.
  • Cloud users utilize organization credits, while self-hosted daemons require a TypeSafe (Jev) API key.

The Bottom Line

Hardcoded routing is a legacy pattern. By abstracting judgment into reusable, testable Decisions, AgentConnect gives builders the control they need to scale agent systems without drowning in conditional logic. This is a significant step toward deterministic yet flexible agent orchestration.