AI summary by TwinMind
How Ramp solved the Fatal Flaw in AI Agent Strategy ft. Rahul Sengottuvelu
Sequoia Capital
5m 52s
Target Audience & Introduction
- Talk for people with distribution - mature companies with users and established software products
- Founders selling to them or investors funding those founders
- Core insight: Don't build tools one-by-one, use computer use approach instead
Common Agent Failure Patterns
- Examples from major companies showing incomplete agent experiences
- Flight booking works but seat changes fail immediately
- Google Slides AI can't bold text ("Sorry, I can't do that")
- Siri limitations ("Sorry, I can't help with that")
- Pattern: feature incomplete, second-class experiences frustrating users
- All companies failing in same way despite being biggest software companies
Technical Architecture Problem
- Simple mental model: frontend → backend
- Reality: mature apps become complex and wide with tons of API endpoints
- Horizontal growth creates messy architecture over time
- Agent teams stuck on long path to feature parity with existing frontend
- Slowly building more tools for LM to use, but always behind
Resource Allocation Issues
- Main product has "jacked frontend team" with PMs, designers, UX people
- Agent team "strapped on to the side" trying to catch up
- Agent team much slower/weaker than main product team
- Starting far behind and staying behind with every new feature release
Traditional Tool Building Approach Problems
- Frontend easy, success depends on tools provided
- Tools also called MCP, ADA, or other terms
- Need tool calling interface over ALL features, not just some
- Authentication complexity - user roles, provisioning, enterprise permissions
- Reinventing wheel instead of leveraging existing frontend work
RAMP's Computer Use Solution
- Core insight: Computer use yourself on frontend instead of building backend tools
- Build heuristics and scaffolding on top of frontend
- Example: Card branding change feature
- Niche feature that wouldn't get tool priority in traditional approach
- Computer use agent spins up browser with user's credentials in background
- Navigates frontend automatically
- User only sees successful result, not the process
Advantages of Computer Use Approach
- Support all tools in one fell swoop vs one-by-one building
- Feature complete agent immediately
- Simplifies task for computer use agent
- Can build own nav tree with DOM heuristics
- Render component library into CLI
- Scaffold unreliable parts and focus on what's broken
- Start with everything vs starting with nothing
