Outline - Building AI Agents In The Browser

Speaker

Lars Trieloff.

Safety Status

This is a safety-redacted public outline. Operational control mechanics and sensitive integration details are omitted.

Thesis

Browser-native agents can become more useful when their UI, context, skills, events, and helper agents are designed as one cohesive experience. That usefulness must be balanced with clear capability boundaries.

Section Map

1. Motivation for an agent close to the browser workflow.
2. Skills as a way to make agents useful in specific contexts.
3. Browser-native agent product model.
4. Generated UI elements and temporary task-specific interfaces.
5. Event-driven interaction between UI and agent.
6. Helper agents and focused delegation.
7. High-level architecture of the agent loop and UI shell.
8. Redacted operational integration details.
9. Safety lessons for browser-native agents.
10. Closing product takeaways.

Concepts

• Browser-native agent: An agent experience that lives close to the user's browser context.
• Skills: Packaged knowledge that helps the agent understand a domain or workflow.
• Generated UI: Task-specific interface fragments created for short-lived work.
• Events: Scoped signals connecting UI state and agent behavior.
• Helper agents: Focused sub-agents that handle narrower work.
• Capability boundary: The line between what the agent can observe, suggest, and do.

Not Included

• Detailed application-control mechanics.
• App-protection bypass details.
• Public event-channel wiring.
• Sensitive-auth handling details.
• Runnable examples from the live demo.

Safe Use

Use this talk to discuss agent UX, browser-centered product architecture, event-driven interfaces, helper-agent design, and capability-scoped safety.