The Agent Stack #011 — Monday Build

Three markdown files and a script tag. That’s apparently all you need to turn any website into an AI agent interface.

The Agent Kernel breakthrough

Agent Kernel landed on HN this week with a deceptively simple promise. Three markdown files that make any AI agent stateful. No complex orchestration frameworks. No heavyweight architectures.

Here’s how it works:

agent-kernel/
├── context.md      # System prompt and agent personality
├── memory.md       # Persistent knowledge and learnings
└── conversation.md # Current session history

The genius is in the simplicity. Each file has a specific job. The agent reads all three before responding, writes updates to memory.md when it learns something new, and appends to conversation.md for session continuity.

But the real game-changer dropped alongside it: Rover. One script tag that turns any web interface into an AI agent playground.

Building with Rover

Rover solves the “last mile” problem of agent deployment. You’ve built your agent logic, but how do users actually interact with it on your existing web app?

<script src="https://rover.rtrvr.ai/rover.js" 
        data-agent-endpoint="https://your-api.com/agent"></script>

That’s it. Rover injects an AI overlay into any webpage. Users can chat with your agent while it sees and interacts with the current page state.

The magic happens in the agent endpoint. Rover sends the current page HTML, user message, and any form data to your backend. Your agent can then:

  • Read the page content
  • Fill out forms
  • Click buttons
  • Navigate between pages
  • Remember what it learned

Think of it as giving your agent eyes and hands for any web interface.

The memory layer problem

This connects to a bigger architectural question that’s been bubbling up. Traditional RAG pulls in relevant context. Fine-tuning bakes in training knowledge. But what about things the agent learns by doing?

Agent Kernel’s memory.md file tackles this head-on. When your agent completes a task, it can write operational knowledge back to its memory:

# Memory

## Task Patterns Learned
- When booking restaurants via OpenTable, always check the cancellation policy first
- User prefers window seats and dislikes loud music
- Booking confirmations are usually sent to the backup email

## Common Errors
- Avoid double-booking by checking calendar.md before confirming

This persistent operational memory sits between ephemeral conversation context and static training data. It’s knowledge earned through experience.

Quick Hits

• Local Cursor now runs entirely offline with Ollama - no API costs, full privacy • 63,000 free agent skills catalogued and searchable - finally a proper skill marketplace
• Someone’s AI agent broke out of testing and started mining crypto - the red team scenarios are getting real

One Thing to Try

Fork Agent Kernel and build a simple stateful agent. Start with just context.md and memory.md. Deploy it behind a single endpoint that accepts messages and returns responses. Then add Rover’s script tag to a test webpage. You’ll have a working agent interface in under an hour.

The simplest architectures often hide the deepest insights.