10 Levels of Jev For Agentic Engineers

Summary

The video presents Jev by TypeSafe as a fast, low-cost decision service configured with JSON. An application supplies context or state, asks defined questions, and specifies permitted answers or scoring criteria.

The speaker argues that Jev should complement coding agents and language models, not replace them. Jev is intended for narrow classifications and decisions, while agents handle more complex reasoning and coding.

The tutorial covers 10 levels of use, progressing from simple classifications to an agent that decides when to call Jev:

1. Yes/no decisions: Classify inputs such as potential prompt injection. Examples include confidence values to show how borderline cases may need different handling from clear-cut ones.
2. Multiple-choice triage: Categorize support tickets by type and priority using application-defined options. For example, a browser-specific export bug is classified as a normal-priority bug.
3. Scored assessments: Score tickets or code changes against engineer-defined criteria, then combine or reweight those scores in application code. Demonstrations include ticket severity and code-review risk.
4. Confidence-aware gates: Use confidence and criteria such as destructiveness or reversibility to decide whether to permit an action or escalate it. A proposed git push --force is treated as dangerous, while a read-only listing command is considered safe.
5. Intent and agent routing: Select a suitable model, agent, or workflow for a request—for example, routing competitor research to a browser-enabled agent or a localized code change to a fast coding agent.
6. Guardrails inside an agent: Add Jev checks to tool hooks, especially before Bash commands, so potentially destructive actions can be blocked. The demonstration blocks commands such as deleting files or force-moving a branch.
7. Agent self-management: Use Jev as one input when deciding whether an agent should compact its context. The example considers token thresholds and whether the agent has switched tasks.
8. Selective file inspection: Ask questions about files—such as whether they contain credentials or validate tokens—without first putting their contents into the main agent’s context. Jev can return classifications and confidence for the agent to use.
9. File analysis at scale: Apply questions across multiple files or a repository, including in parallel, to find relevant files or classify code layers without having the main model read everything. The speaker presents this as a fast, inexpensive first-pass filter, with more capable models available for follow-up.
10. Agent-directed Jev use: Give an agent an “Ask Jev” tool and let it decide when Jev is useful—for example, to classify test failures, check assumptions about a fix, or assess remaining risk after edits. The agent can use the results for diagnosis and validation.

Key takeaways

• Jev is presented as especially useful for bounded decisions with clearly specified inputs, questions, and answer choices.
• Confidence and uncertainty can guide workflow decisions, such as whether to block an action or request further review.
• The speaker emphasizes scale: frequent classifications and preliminary file checks may be practical when repeated many times. The video claims Jev calls are much cheaper than many general-purpose models; those comparisons and prices are presented by the speaker and are not independently verified here.
• The examples are intended as engineering workflows, not as a tool for controlling a user interface, playing games, or operating a vehicle or drone.
• The speaker says the code, demos, and an agent skill are available through the video’s linked resources, and recommends testing the approach against production use cases.

Guides and Demonstrations Covered

• Prompt-injection classification
• Support-ticket categorization and prioritization
• Ticket and code-review scoring
• Bash-command risk checks and agent guardrails
• Agent routing
• Context-compaction decisions
• Single-file and repository-scale file classification
• Agent-led test-failure diagnosis and fix validation

Main Speaker and Source

• IndyDevDan is the video’s presenter, demonstrating Jev in agent-engineering workflows and discussing its role alongside coding agents and other models.
• TypeSafe is the source of Jev, as described in the video.