The interesting problem is not clicking buttons. It is adapting to messy workflows, grounding actions in real user context, learning from failure, and knowing when to ask for review.
This project explores a browser agent that can help with high-friction application flows, without automating away the human judgment that actually matters.
What I'm building
- Understanding recurring web workflow patterns.
- Drafting form answers from approved source material.
- Remembering where workflows break.
- Pausing for human review before any sensitive or final action.
This is not a mass application bot, a recruiter bypass system, or anything that submits without explicit approval.
Open questions
- How should the agent represent form state and user intent?
- What failures should be remembered for the next attempt?
- Where should the system pause and ask for review instead of guessing?