Lee Deep Dive: OpenClaw Architecture, Agent Teams & HC Protocol

Feb 10, 2026 28 min Jason + Lee Gunderson
Wide-ranging technical discussion: Forge multi-model router architecture, Claude Code Agent Teams vs GSD orchestration, Lee's GitHub Copilot experience at Best Buy (code reviews 1hr → 5min), and a full HC protocol walkthrough.

Forge Router Architecture

Jason's proposed multi-model escalation system for the Forge "Jarvis" assistant:

Router
Local Qwen (free)
OpenRouter (Kimi/Qwen large)
DeepSeek (local only)
PII Guard
Claude Haiku

Design Principles

  • Start cheap — local models handle bulk of simple queries
  • Ralph Loops: if answer quality fails, escalate to next tier
  • PII scanning layer (Microsoft LiteLLM) before any cloud model
  • DeepSeek: local-only, never Chinese-hosted versions
  • Budget + time stops to prevent token runaway
  • Docker containers: persistent for core services, ephemeral for tasks
  • Git backup before every destructive action
  • Standalone email/calendar for agent (not personal accounts)

Discussion Topics

Claude Code Agent Teams vs GSD

Lee discovered Agent Teams (experimental, JSON settings toggle). Compared to GSD orchestrator — both coordinate multi-agent work, but Agent Teams is first-party with inter-agent communication. Key insight: orchestrator/lead agent uses lots of tokens, so Agent Teams best for complex projects, sub-agents for simpler tasks.

"Before, you always had to plug stuff together. This is the first step in what it's gonna start looking like." — Lee

GitHub Copilot at Best Buy

Lee's team using Copilot's coding agent at Best Buy. One team cut a 4-month project to 1.5 months. Lee uses Copilot code reviews — reviews that took 1 hour now take 5 minutes. Training offshore devs to self-review with AI before submitting.

"Not everything they come up with is legitimate. I just take off the ones that don't matter and send it back." — Lee on AI code review

HC Protocol Walkthrough

Jason demoed NowPage publisher and explained the HC Standard concept:

  • Pages as knowledge base — HTML with agentic metadata, not just SEO metadata
  • Lazy-load architecture: header exposed first, context loaded on demand
  • Registry = dynamic data source (add pages to registry, agent discovers them)
  • Injection: per-user registry (brand, persona, writing style) customizes shared skills
  • Executable web pages — not just read-and-paste, but agent-actionable
  • Like N8N workflows but using World Wide Web + URLs

Steelman / Recursive Thinking

Jason's "two-hat" recursive approach predates Agent Teams' devil's advocate pattern. Instead of destroying bad ideas, the steelman forces the solution to be stronger. Links to emerging Recursive Language Models (RLMs) research.

Traditional Dev vs AI Coding

Jason arguing traditional programming is "dead" — built entire NowPage infrastructure without coding knowledge. But skilled devs still add value by directing AI more specifically (CSS cleanup, architecture choices). Proposed race: same task, manual dev team vs Claude Code.

Key Takeaways

Next Steps