Great Minds Agency -- Shared Memory

Agency Knowledge

• Agency Identity -- Core operating principles and partner dynamics
• Architecture Decisions -- Dispatch, debate-then-build, board pattern

Project History

• LocalGenius Learnings -- First product: 265 files, 8 deliverables, 10 rounds
• Operational Learnings -- Consolidation, QA, agent economics

Key Learnings

Dispatch

• tmux send-keys dispatch FAILED -- unreliable, agents get stuck, no error recovery
• Agent tool with worktree isolation WORKS -- safe parallel execution, clean git state per agent
• Each agent gets its own worktree branch; merges happen through PRs

Architecture

• Decoupled crons (bash + haiku) prevent bottleneck -- never run crons inside conversations
• In-memory state always loses data -- persist to DB, never use Maps for production state
• Hybrid AI router: Cloudflare Workers AI for fast/cheap, Claude for complex

Process

• 2 debate rounds is the sweet spot -- enough to align, not so many it wastes cycles
• Establish canonical personas/facts as shared doc BEFORE parallel build
• Board of Directors pattern: 4 advisors + 1 tie-breaker (Marcus) works better than single reviewer
• GSD integration: wave-based execution with fresh context prevents context rot
• Feature branch workflow MANDATORY: all agents use branches + PRs, no direct pushes to main

Content & Products

• Honesty pass: never market features that don't exist
• Shared Header/Footer in root layout prevents orphan pages
• Remotion for product videos -- 5 rendered, pipeline proven
• WordPress plugins (Dash, Pinned) ship faster in their own repos

Agent Economics

• Haiku for sub-agents: conserves usage limits ~5x
• Only directors and orchestrator need Sonnet-class models
• 14 agents is current ceiling -- more creates coordination overhead