

Flowing with agents (Interview)
67 snips Sep 17, 2025
Join Beyang Liu, CTO at Sourcegraph and a leader in code intelligence and developer tools, as he dives into the revolutionary world of Amp. He explains Amp's multi-model approach that enhances coding agents, discussing its unique architecture and capabilities. Discover insights on Adam's "Agent Flow" concept, which facilitates long-term project management. They explore best practices for agent workflows, logging, and the importance of community learnings in shaping Amp's future. If you're keen on maximizing coding with AI, this chat is a must-listen!
AI Snips
Chapters
Transcript
Episode notes
Models As An Implementation Detail
- AMP treats model choice as an implementation detail and composes multiple models for different sub-capabilities.
- Users see a single agentic experience without toggling specific LLMs.
Treat Prompting As A Learned Skill
- Learn prompting as a skill and give agents sufficient context to avoid poor outputs.
- Provide clear instructions and more tokens for complex tasks to reduce guesswork.
Never Skip Human Code Review
- Always review and code-review agent outputs; agents are powerful but not fully trustworthy.
- Read generated code to catch subtle bugs and prevent slop from creeping in.