# Why Are People Excited About Jev?

Erica Hughberg | Debug Mode | 1:04:37
Hosted by Demetrios Brinkmann

Source: https://www.youtube.com/watch?v=QbZkd3SVO3U
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/why-are-people-excited-about-jev
Published: 2026-11-06
Tags: agents, mcp, security, tool-use

## TL;DR
- Skills over MCP could make skills easier to discover and keep updated, but loading more skills can reduce an agent's performance.
- Jev does not introduce classification itself. Its appeal is a fast, LLM-native API that makes classification easier to add to products.
- Computer use is likely to remain useful because people need interfaces for collaboration, visual expression, and tasks that never receive an API or MCP server.

## Summary
Demetrios Brinkmann and Erica discuss skills over MCP, Jev, agent security, and the future of computer use. Skills over MCP could let vendors distribute current skills without users maintaining them, but the combination of models, services, and hidden skills makes evaluation difficult. They see Jev as an accessible, fast classifier API rather than a new machine-learning idea. It could decide when to involve a human in customer support, prefetch information for a voice agent, or check whether a child's question is still about maths. The conversation then turns to WebMCP and authorization. Erica argues that OAuth was designed to stop people sharing passwords with third parties, so it is a poor cage for more capable agents. The hosts compare this to putting a bear in a bunny cage. They expect computer use to remain alongside MCP and APIs because websites support human participation, visual design, and tasks that services may never expose headlessly.

## Key ideas
### Skills over MCP could solve the problem of keeping distributed skills current
[08:18](https://www.youtube.com/watch?v=QbZkd3SVO3U&t=498s)
Erica describes trying to consume skills over MCP while ensuring that every consumer gets the latest version. Skills over MCP could act as a distribution mechanism, allowing vendors such as AWS, Microsoft, or Google to publish skills for their services. Users could discover and subscribe to skills instead of finding and updating them manually. The discussion also raises a harder problem: a provider may need to support many services, models, and model-specific skill implementations. A single large MCP server may be difficult to manage, while thousands of small servers create their own maintenance burden.

### Adding more skills can make an agent perform worse
[12:05](https://www.youtube.com/watch?v=QbZkd3SVO3U&t=725s)
Erica explains that skills do not automatically improve an agent when added in larger numbers. Different skills can interact badly, and the right set depends on the task, the model, and other skills already loaded into the agent. Rick is working on an evaluation framework to test skill combinations and find sets that improve results without making outputs worse. A vendor cannot know which model a user has chosen or which hidden skills are already present. Users therefore depend on providers to create skills that work well in situations the provider cannot fully see.

### Jev is exciting because it makes classification easy to use at speed
[16:36](https://www.youtube.com/watch?v=QbZkd3SVO3U&t=996s)
The hosts agree that classifiers are not new. Erica recalls seeing classifiers used in machine learning for years and says Jev can still hallucinate or produce a misleading confidence score. What feels new is the interface. A developer can provide context, a question, and the allowed answer types through an API without building infrastructure or training a specialised classifier. Demetrios describes this as an LLM-native way to use classification, since developers already understand prompts and context windows. Erica expects best practices for structuring Jev requests to emerge over time.

### Fast classification can decide when an agent should involve a human
[20:48](https://www.youtube.com/watch?v=QbZkd3SVO3U&t=1248s)
Erica suggests putting Jev on every turn of an AI customer assistant to decide whether a human should join before the customer becomes frustrated. The system could recognise repeated questions or requests that the assistant cannot handle, rather than waiting for the agent to fail. Demetrios adds a voice-agent example. Jev could classify the direction of a conversation and send a worker to fetch information before the human asks for it. The worker's result would then be ready when the voice interaction reaches that point.

### A maths tutor could use Jev as a quick topic check for a child
[24:13](https://www.youtube.com/watch?v=QbZkd3SVO3U&t=1453s)
Erica describes a tutor for her seven-year-old niece, who does not enjoy maths. She wanted the tutor to keep the conversation about maths and respond appropriately for a young child, but slower guardrail approaches could make the interaction frustrating. Jev could quickly assess whether a question is about maths and return a confidence level before the main response is generated. Erica does not present this as a replacement for guardrails. She sees it as a small, fast service that can handle one narrow decision before the rest of the system runs.

### OAuth and existing safety boundaries were built for older systems
[33:31](https://www.youtube.com/watch?v=QbZkd3SVO3U&t=2011s)
Erica explains that OAuth was created to stop people giving passwords to third-party applications. It later adapted to the API era, but agents introduce a different problem because they can operate across several layers and services. The hosts say MCP providers and consumers need to understand the authorization and access decisions they are making instead of assuming the standard solves them automatically. Erica compares current protections to a bunny cage made from chicken wire. It may work for a rabbit, but a bear can break through it. Her concern is that people are upgrading agents faster than they are upgrading the cages around them.

### Computer use will remain alongside MCP and APIs
[43:52](https://www.youtube.com/watch?v=QbZkd3SVO3U&t=2632s)
The hosts question whether computer use is only a temporary bridge until every website becomes headless and exposes MCP or APIs. Erica expects computer use to remain because some organisations, such as government services, may never provide the needed MCP server. She also compares computer use to a cobot, where a machine works alongside a person rather than replacing the person. Demetrios adds that people are often bad at explaining what they want in chat, while showing an agent what to do can communicate intent directly. Websites also allow people to experience visual design and interact with interfaces in ways a plain chat window cannot reproduce.

### Chat-only interfaces could lose the expressive parts of the web
[52:03](https://www.youtube.com/watch?v=QbZkd3SVO3U&t=3123s)
The discussion moves from agent tooling to design and social interaction. Erica enjoys watching computer use happen and describes the appeal of seeing an agent work through a browser. She and Demetrios also talk about websites with distinctive design, hidden Easter eggs, and carefully made user experiences. Erica used the Life of a Showgirl colour scheme from Taylor Swift's work for an open-source project, as a private design reference. They compare this kind of discovery with Stumble Upon and CSS-focused web galleries. A chat window can make digital interaction feel like plain text files, and the hosts do not think people are ready to give up the visual and social parts of the web.

## Notable quotes
- Erica: "I feel the problem is not the agent, it's we are bad at building cages and the safety mechanisms." (41:51)
- Erica: "I think the thing people aren't actually praising Jev for is ease of use. It's developer experience." (18:38)
- Demetrios Brinkmann: "I think it's like bringing that design to the masses." (28:43)
- Erica: "I do not believe we're going to go fully headless because going fully headless would be working on something that isn't necessary." (46:10)
- Erica: "Dealing with just the chat feels very boring. Just march down files. Like really, that's the best we've got. We've gone backwards." (1:00:32)

## Tools & references mentioned
- Ola
- Skill MCP
- Model Context Protocol
- AWS
- Astra
- GCP
- Microsoft
- Google
- Amazon
- James Ward
- Jev
- Codex
- Claude
- Rick
- BERT
- Anthropic
- OpenAI
- WebMCP
- OAuth
- MCP apps
- LinkedIn
- Gmail
- Tesla
- Taylor Swift
- Life of a Showgirl
- Stumble Upon
- Facebook
- Microsoft Excel
- Gates Foundation

## Who should watch
- You are deciding whether to add skills, tool search, or fast classifiers to an agent and need to understand the failure modes of combining them.
- You are building customer-support, voice, or educational agents where a quick decision can determine whether to call a human or continue.
- You are designing authorization, browser automation, or WebMCP integrations and want a practical argument for keeping computer use in the system.

## Related talks

- [MCP is not going to change everything (yet)](https://mlopstalks.com/talks/mcp-is-not-going-to-change-everything-yet) (Sam Partee, Arcade AI & Rahul Parundekar, AI Hero, 1:04:43)
- [From AI Slop to Software Factories](https://mlopstalks.com/talks/from-ai-slop-to-software-factories) (Dexter Horthy, HumanLayer, 47:01)
- [AI Agents Need a Babysitter](https://mlopstalks.com/talks/ai-agents-need-a-babysitter) (, 45:53)
- [Enterprise-ready MCP](https://mlopstalks.com/talks/enterprise-ready-mcp) (Jiquan Ngiam, MintMCP, 28:36)
- [AI Agents Are About to Get Real, Here's Why It Matters](https://mlopstalks.com/talks/ai-agents-are-about-to-get-real-heres-why-it-matters) (Sam Partee, Arcade AI, 13:20)
