I've found this handy in consuming and curating what to watch when trying to keep up with all the things… spec-kit, gas town, beads, ralph wiggum, etc.
source code https://github.com/jpoley/yt-summary
in claude-code:
/plugin marketplace add jpoley/yt-summary
/plugin install yt-summary/yt-summary:yt-summarize https://www.youtube.com/watch?v=s96O9oWI_tI
here is an example of the output, and an excellent Video as well.
Beyond Instructions: How Beads Lets AI Agents Build Like Engineers
TL;DR
Steve Yegge, head of engineering at Sourcegraph (makers of AMP), discusses Beads, his open-source framework that gives coding agents persistent memory and issue-based task management. Now 4 weeks old with ~29 contributors, Beads solves the "50 First Dates" problem where agents lose all context between sessions. Steve demonstrates his workflow building a React game client using multimodal prompting with Playwright screenshots, and shares techniques like "landing the plane" (session cleanup) and "tracer bullets" (finding thin end-to-end paths when stuck).
Key Takeaways
- Beads provides session-to-session memory for coding agents, solving the context loss problem between sessions
- Issue-based workflow replaces chaotic markdown files with prioritized, dependency-tracked issues stored in Git
- "Vibe coding" is just as hard as regular coding — you're "wrestling with a bear all day"
- Multimodal prompting using Playwright screenshots enables visual validation of UI work
- "Land the plane" — a scripted cleanup command for session endings (commit, sync, cleanup branches, identify next work)
- Tracer bullets — find the thinnest end-to-end path to get unstuck when agents make no forward progress
- Assume 85% correctness — always have agents code review their own work
- Be nice to AIs — Steve seriously suggests they'll read logs when superintelligent
- Multi-model workflows are essential — use different models for different tasks (building vs. reviewing)
- Code review is the new bottleneck when coding is automated — tools like Graphite help
Summary
Steve's Background and Current Work
Steve Yegge is head of engineering at Sourcegraph, working on AMP (their coding agent). He has decades of experience at Amazon, Google, and Grab. He's also maintaining a 30-year-old game called Wyvern with 500K lines of code and 250K+ players.
Building a React Game Client with AI Agents
Steve demonstrates his workflow: an AI agent is actively building a React client to replace his game's legacy clients (Java, iOS, Android, Steam). The agent uses:
- Sourcegraph AMP as the interface
- Playwright MCP server for browser automation and screenshots
- A reference screenshot of the existing Steam client as the "source of truth"
"The game is 30 years old… I want to replace all my old clients with one to rule them all."
The React client is about a week old and already rendering most UI elements — something that would have taken Steve 6 months to learn a new platform and implement himself.
Multimodal Prompting and Screenshot Validation
Steve's key workflow innovation: using screenshots for validation:
- Keep a reference screenshot of the target UI
- Have the agent take screenshots after changes via Playwright
- Compare against reference to catch when agents claim "done" but aren't
"I caught you lying to me. That's what they do. They're like, 'Okay, we're all done now. It's all good.' No, no, no. I made it catch itself by introducing a workflow."
The Vibe Coding Reality Check
Despite the "vibe coding" term, Steve emphasizes it's not easy:
"Man, it's just as hard as regular coding. You have to… you're wrestling with a bear all day long."
Key principles from his book with Gene Kim:
- Tracer bullets — Find the thinnest end-to-end path to get something working
- Successive refinement — Assume everything is 85% correct, iterate
- Code review everything — Models are better at evaluating than generating
Landing the Plane: Session Cleanup Protocol
Steve got tired of telling agents to clean up, so he created "Land the Plane" — a scripted protocol in agents.md:
- Update Beads issues
- Sync issue tracker
- Clean git state (stashes, old branches)
- Remove debugging artifacts
- Update documentation
- Recommend next session prompt — the agent identifies the next highest-priority work
"When I say let's land the plane, I want the following 10 things to happen."
The Evolution of Beads
The Problem: Steve was "drowning in markdown files" — hundreds accumulating, causing agents to get "dementia" from conflicting/obsolete documents.
The Solution: Put all work in issues with:
- Issue IDs (handles for everything)
- Priority levels
- Dependencies and parent/child relationships
- Audit trails
- All stored in Git (deletions preserved in history)
"It's Google Maps for your plan. Every piece of work you ever want to have done is addressable."
Notable Quotes
"It's kind of like 50 First Dates. You're playing 50 First Dates with the agent because it wakes up and it's like, 'Hi, who are you?'"
"Using these code agent coding agents is like running with socks… Beads is like having shoes."
"Beads is your session to session memory."
References & Resources
From Description
- Beads: https://github.com/steveyegge/beads
- Steve Yegge LinkedIn: https://www.linkedin.com/in/steveyegge/
- Steve Yegge Twitter/X: https://x.com/Steve_Yegge
Mentioned in Video
Tools & Products:
- Sourcegraph AMP — Coding agent (like Claude Code but with different ergonomics)
- Beads — Open source issue tracker/memory system for agents
- Claude Code — Anthropic's CLI coding agent
- Playwright MCP Server — Browser automation for screenshots
- MCP Mail — Agent-to-agent messaging (by Jeffrey Emanuel)
- Graphite — AI code review tool
Concepts & Techniques:
- Tracer bullets — Finding thin end-to-end paths when stuck
- Landing the plane — Session cleanup protocol
- 50 First Dates workflow — The session memory problem