AI agent stack

Multi-model orchestration. One orchestrator, many models, escalation lanes, and a multi-vendor worker fleet behind it. Most of these guides assume OpenClaw as the agent framework, but the patterns generalize. The worker layer (Brigade roster seats and T3 Code delegated tasks) is described in multi-model-orchestration.md and sub-agent-patterns.md.

Browser Chromium lane

Not every useful model path is an API.

This stack also uses a dedicated Chromium lane for browser-native AI workflows where logged-in sessions matter more than clean API abstractions. That includes things like:

• Perplexity research from a logged-in Pro account
• Gemini browser workflows
• Claude web workflows
• ChatGPT web workflows
• any task where cookies, subscriptions, uploads, or persistent session state are the real dependency

The pattern is covered in browser-llm-stack.md:

• keep a persistent Chromium profile for each provider or workflow
• run it in a controlled environment
• serialize fragile browser tasks with a lock when needed
• treat browser automation as a first-class lane, not an embarrassing fallback

This matters because some of the highest-value agent workflows still live behind browser auth, subscription entitlements, or UI-only features. If your stack ignores that reality, you end up designing around an imaginary clean-room API world.

🦞 Per-guide format lives in ../automation/cron-patterns.md.