How Jev by TypeSafe changes agentic engineering?

TL;DR

Jev by TypeSafe is programmable through JSON for intelligent question answering that can be scaled to production. It supports levels of decision making from simple yes no to complex composite scoring and confidence gating, enabling engineers to offload hard calls to a system one model while keeping gates and human review where needed.

Transcript

What's up engineers? Indie Devdan here. Jev by Typesafe is materially changing the way I think about building with agents. You'll see exactly what I mean in this video. By now you've heard of Jev by Typesafe. Let's strip away all the hype and answer what is Jeb really. Jev is intelligent question answering that's programmable through JSON. Instead ... Read More

Key Insights

• Jev is intelligent question answering programmable through JSON.
• Jev pricing emphasizes scale, enabling millions of executions at low cost.
• Level 1 is a basic yes or no decision with confidence intervals.
• Level 2 adds multiple choice options for triage tasks.
• Level 3 introduces composite scoring with weighted inputs.
• Level 4 introduces confidence gating to block dangerous or irreversible actions.
• Composite scoring uses defined weights rather than prompts for easy tuning.
• The core message is prompt engineering remains crucial for effective model communication.

Questions & Answers

Q: How does Jev function at its basic level and what problem does it solve?

Jev at level one functions as a basic yes or no decision tool that operates with a confidence interval. It is designed to classify inputs such as prompt injections, determining when a simple decision is safe to execute. This level aims to provide a fast, cheap decision mechanism suitable for integration into agentic workflows.

Q: What is the role of confidence intervals in Jev level one?

In level one, the confidence interval quantifies how certain Jev is about its yes or no classification. The interval guides whether the decision can be acted on immediately or should be treated with caution. This helps engineers decide when a human review or additional checks are necessary.

Q: How does level two improve Jev for triage tasks?

Level two adds multiple choice options to Jev so you can select from a defined set of outcomes. This makes it suitable for triaging issues such as bug reports or priority setting, enabling faster categorization while keeping responses aligned with predefined categories.

Q: What is composite scoring in Jev level three and why is it useful?

Composite scoring in level three aggregates multiple inputs into a single score using defined weights. This allows nuanced assessments like ticket severity or code review risk to be scored consistently, with adjustments made simply by changing weights rather than prompts.

Q: What is confidence gating in level four and when is it used?

Confidence gating uses the model score to decide whether to proceed with an action, such as executing a command. It helps prevent irreversible or high risk actions by ensuring a gate is crossed only when Jev is sufficiently confident in the outcome.

Q: How does Jev demonstrate cost advantages over other models?

The video highlights that Jev can perform millions of executions at low cost, comparing its economics to larger, more expensive models. This affordability enables production-scale automation where quick, repeated judgments are needed without breaking the budget.

Q: What do the levels suggest about the future use of Jev in engineering workflows?

The levels illustrate a progression from basic decision making to sophisticated gating and scoring, suggesting Jev can handle increasing responsibility in production systems. This enables engineers to delegate routine calls to Jev while reserving critical decisions for human review or more capable components.

Q: What is the overall message about prompt engineering in relation to Jev?

The speaker stresses that prompt engineering remains essential, describing it as a core skill for communicating with system one models and classic LLMs. Even with Jev’s structured approach, effective prompts and well designed input schemas are crucial to achieve reliable, desirable outcomes.

Summary & Key Takeaways

• Jev is introduced as a programmable intelligent question answering system that uses JSON inputs to drive decisions across an engineering workflow.

• The video outlines 10 levels of Jev, starting with basic yes no decisions and advancing to confidence gated and composite scoring scenarios for production use.

• The speaker emphasizes cost efficiency and scale, showing how Jev enables rapid, cheap executions suitable for production environments without overreliance on large language models.