Jev, Clearly Explained
Beacon, developed by Asymptote Labs, is an open-source telemetry and memory layer designed to capture and share knowledge across different AI coding agents. It addresses the issue of agent amnesia by extracting reusable workflows and debugging patterns from agent traces. By integrating with Jev for cost-effective evaluation, Beacon identifies high-quality runs to build a persistent knowledge base. This allows improvements made in one tool, like Cursor, to benefit others, such as Claude Code or Codex, making agent-derived knowledge portable across an organization. TypeSafe AI has launched Jev, a specialized semantic decision engine designed to replace generative LLMs for bounded tasks like routing, risk assessment, and classification. Unlike traditional LLMs, Jev does not generate text but instead provides typed outputs—Choice, Score, and Noul—along with calibrated probabilities. This approach reduces latency to sub-500ms levels and significantly lowers costs by eliminating token generation and parsing. Jev is intended to work alongside generative models, handling the discrete judgment calls within agentic loops and software control flows.
閱讀原文 ↗目錄
Jev makes agent evaluation cheap. Beacon turns it into cross-harness memory
Beacon, developed by Asymptote Labs, is an open-source telemetry and memory layer designed to capture and share knowledge across different AI coding agents. It addresses the issue of agent amnesia by extracting reusable workflows and debugging patterns from agent traces. By integrating with Jev for cost-effective evaluation, Beacon identifies high-quality runs to build a persistent knowledge base. This allows improvements made in one tool, like Cursor, to benefit others, such as Claude Code or Codex, making agent-derived knowledge portable across an organization.
- Beacon is an open-source tool that acts as a cross-harness memory layer for AI coding agents.
- Jev provides a low-cost method for evaluating agent traces to identify high-signal runs.
- Beacon supports over 19 agent harnesses, including Claude Code, Cursor, Codex, OpenCode, and Cline.
- The system converts successful agent traces into reusable skills and workflows for future sessions.
- Knowledge captured by Beacon is portable, meaning learning from one agent can be applied to a different agent.
- The workflow involves capturing traces, evaluating them with Jev, and extracting knowledge to improve future agent performance.
Jev, clearly explained:
TypeSafe AI has launched Jev, a specialized semantic decision engine designed to replace generative LLMs for bounded tasks like routing, risk assessment, and classification. Unlike traditional LLMs, Jev does not generate text but instead provides typed outputs—Choice, Score, and Noul—along with calibrated probabilities. This approach reduces latency to sub-500ms levels and significantly lowers costs by eliminating token generation and parsing. Jev is intended to work alongside generative models, handling the discrete judgment calls within agentic loops and software control flows.
- Jev is a non-generative semantic engine that outputs structured data and probabilities instead of prose or code.
- The tool supports three specific output primitives: Choice (categorical), Score (ordinal), and Noul (Boolean probability).
- It is trained using Reinforcement Learning for Calibrated Decisions (RLCD) to ensure confidence scores match observed accuracy.
- Reported performance includes latency between 70 and 500 milliseconds and a cost of $0.042 per million input tokens.
- LangChain has integrated Jev to provide middleware for safety checks and agent gating.
- Jev is designed for closed-set outcomes where meaning-based judgment is required but deterministic code is insufficient.
- The engine cannot perform arithmetic, extract unknown values, or explain its reasoning, making it a complement to, rather than a replacement for, LLMs.