# Why Are People Excited About Jev?

## Executive summary

The discussion centers on the rapid advancement of AI agents and the critical need for robust infrastructure and security mechanisms to contain them. The core argument is that the problem is not the agents themselves, but the outdated 'cages' (security and authorization frameworks) built around them. Key technical areas covered include the evolution of Multi-Capability Platforms (MCP) and WebMCP for exposing complex website user journeys as tools, the use of fast classification APIs (Jev) for real-time decision-making (e.g., human handoffs, tool selection), and the inadequacy of traditional authorization methods like OAuth for the new agent paradigm.

## Key takeaways

- The Agent Security Problem (Bear in a Bunny Cage): The analogy suggests that current security measures (like OAuth) are insufficient for powerful AI agents. These systems are designed for human-to-human or application-to-application interaction, not for containing a powerful, autonomous agent. The problem lies in the cage, not the animal.
- WebMCP as the Next Frontier for Agent Interaction: WebMCP is viewed as a natural evolution for allowing agents to interact with the full complexity of websites. Instead of exposing millions of individual tools, the focus should be on exposing the top four 'core user journeys' or 'happy paths' of a website, making the agent experience more manageable and robust.
- Jev's Value: Developer Experience and Speed: While classification itself is not new, the excitement around Jev (the fast classifier API) stems from its developer experience. It allows developers to easily integrate classification into their system via a simple API call, enabling fast, real-time decisions (e.g., determining if a customer needs a human handoff or if a math question is still on topic).
- The Limitations of Skills and Tool Selection: Loading multiple skills into an agent can sometimes degrade performance. There is a complex challenge in optimizing skill combinations, especially when dealing with large vendor ecosystems (like AWS services), requiring advanced evaluation frameworks to prevent negative output.

## Technical details

- WebMCP/MCP Servers: MCP servers are designed to allow agents to consume capabilities from various services. WebMCP extends this by enabling the exposure of entire website user journeys (core flows) as structured tools, moving beyond simple API calls.
- Jev (Fast Classifier API): Jev provides a fast, LLM-native way to classify input (e.g., user query, conversation turn) with a confidence score. Use cases include proactive human handoffs in customer service or quickly assessing if a conversation remains within a specific domain (e.g., math tutoring).
- Authorization and Access Control: The speaker notes that OAuth, while the most adopted authorization system, was designed for a pre-agent era (stopping people from handing over passwords). It is insufficient for the complex, multi-layered interactions of autonomous agents, requiring a fundamental rethink of the security framework.

## Practical implications

- Build engineers must prioritize developing advanced guardrails and sandboxing mechanisms that are robust enough for autonomous agents, moving beyond simple API wrappers.
- Focus on defining and exposing the 'core user flows' of a system (WebMCP approach) rather than attempting to catalog every single possible function.
- Integrate fast, real-time classification APIs (like Jev) into the agent workflow to manage state, detect topic drift, and proactively escalate to human intervention.
- Re-evaluate existing authorization models (OAuth) to ensure they can handle the non-deterministic, multi-step nature of agent interactions.

## Topics

AI Agents, LLM Architecture, System Security, Web Development, API Design, Jev, MCP (Multi-Capability Platform), WebMCP, OAuth

Source: https://www.youtube.com/watch?v=QbZkd3SVO3U
