PowerUp Coaching — Living Knowledge Base

Session: TrelloAgents — Two Architecture Patterns Worth Stealing

3 min read

TrelloAgents — Two Architecture Patterns Worth Stealing

Type: Build session (Lou solo — no AIMM cohort present) Output: TrelloAgents v1.0 distribution package released to AIMM members

What Happened

Completed and released TrelloAgents v1.0 — a production-ready multi-agent pipeline that transforms product ideas into PRDs autonomously. The system uses a 6-agent team orchestrated by Claude Code, with a Trello board as the shared state machine and dashboard.

Deliverables:

  • Full dist/ package for AIMM member installation (Python scripts, agent prompts, docs, config templates)
  • Teaching article published to AIMM Notion: “I Built a 6-Agent AI Team That Writes PRDs. Here’s Every Design Decision.”
  • 5 SVG architecture diagrams embedded in the article
  • Final production lint — all references, model names, dependencies corrected

Teaching article (Notion): https://www.notion.so/34fbe0e844bb812b839df599351872d0 Distribution package: /Volumes/Extreme Pro/users/loudalo/GitHub/code experiments/TrelloAgents/dist/

Eight Design Patterns Documented

The build surfaced eight reusable multi-agent design patterns:

  1. Kanban as State Machine — Trello columns = states, cards = work items, movement = transitions. Don’t build invisible infrastructure when a visual tool is already a legible state store.
  2. Orchestrator Separation — Claude Code is the orchestrator. It reads board state, delegates to specialist agents, attaches outputs, moves cards. Don’t build custom orchestration infrastructure when an intelligent agent can already coordinate.
  3. Division of Cognitive Labor — Six specialized agents, each knowing only its job. The Spec Reviewer is credible because it didn’t write the spec. Specialization produces quality and clear accountability.
  4. 1 → N → 1 Pipeline Pattern — One input fans out to N parallel workstreams, converges to one output. Parallelism is where the time savings live.
  5. Structured Object as Agent Memory — See Insight - The Structured Object as Agent Memory — Agents Don’t Talk to Each Other, They Talk Through the Object.
  6. Autonomous Loop + Termination Condition — The review-bounce loop runs until quality clears. An iteration counter in card metadata is the termination guardrail. Every autonomous loop needs a termination condition as a first-class design requirement.
  7. Prompt-as-Configuration — See Insight - Prompt-as-Configuration — The Behavior Layer Is Just Text.
  8. Engine vs. Application — The 6-agent pipeline is not a PRD machine — it’s a pattern (one input, N parallel workstreams, convergent output, quality loop). Swap the agent prompt files and it handles content marketing, course curricula, sales proposals, due diligence reports.

Workflow Discoveries

  • Notion MCP disconnects between sessions: Tools appear in deferred list and ToolSearch returns schema, but fail at call time. Workaround: delegate Notion ops to a sub-agent — sub-agents get fresh MCP access. (Now documented in ~/.claude/CLAUDE.md Known Workarounds.)
  • SVG → Notion image workflow: Write SVGs → publish with here-now skill → get public URLs → insert with update_content using ![alt](url). Anonymous here.now sites expire in 24h — claim immediately. (Also documented in CLAUDE.md.)

Derived Artifacts

Graph View