Four Directions Orchestration Pattern

How multi-agent spawning maps to the Medicine Wheel directions in a PDE session.

The Pattern

``` 🌅 EAST — VISION PDE Decomposition (rise-pde-session) | SLEEP 1500s human + companion | v ┌───────── WAKE ──────────┐ | | 🔥 SOUTH 🌊 WEST Deep-Search Validation (parallel agents) (parallel agent) per PDE facet cross-reference | | └──────── MERGE ──────────┘ | v ❄️ NORTH Execution (parallel agents) file-ownership | v INTEGRATION lead agent commit + checkpoint ```

Why Four Directions, Not Pipeline

Traditional CI/CD is a pipeline: build → test → deploy. Sequential. Mechanical.

The Four Directions pattern is a cycle: vision → analysis → reflection → action. Each direction feeds the next. SOUTH findings inform WEST validation criteria. WEST validation shapes NORTH execution constraints. NORTH actions complete what EAST envisioned.

In multi-agent terms:

• SOUTH and WEST agents can run in parallel because they serve different purposes
• NORTH agents run after SOUTH completes (they need the findings)
• WEST validation runs alongside NORTH (preparing the checklist while executors build)

Agent Spawning Rules

From deep-research: What We Learned

The deep-research skill proved these orchestration principles:

1. MECE decomposition — each agent covers a distinct angle with explicit boundaries
2. "Do NOT cover X" — every agent prompt states what OTHER agents handle
3. Batch results before synthesis — wait for ALL agents, then merge. Never process one-at-a-time (anchoring bias)
4. Scale to complexity — 2 agents for simple, 15 for multi-repo ceremony
5. File ownership — no two executor agents touch the same file
6. Lead agent integrates — sub-agents report, lead merges and commits

PDE-Specific Additions

7. Direction-typed agents — each agent knows its Medicine Wheel direction
8. .pde/ as workspace — SOUTH agents write to .pde/<uuid>/deep-search/, never source
9. Rispecs awareness — agents that author rispecs read the survey first
10. Creative orientation throughout — no problem-solving language in any agent output

File Ownership Protocol

``` SOUTH agents → write to .pde/ only WEST agents → write to .pde/ only NORTH agents → write to assigned source files only (exclusive ownership) Lead agent → reads everything, commits everything ```

When assigning NORTH agents:

1. List all files to be created/modified from the action stack
2. Group by natural clusters (all rispecs together, all implementation together)
3. Verify no file appears in two agents' lists
4. If a file must be touched by two action items, batch those items into one agent

Synthesis Protocol

After all agents return:

```

1. Read SOUTH findings → understand what exists, what the framework says
2. Read WEST checklist → know what to validate
3. Read NORTH outputs → see what was created
4. Cross-reference → do NORTH outputs satisfy WEST criteria?
5. Resolve conflicts → if two agents disagree, structural tension analysis
6. Integration test → build/lint if applicable
7. Commit → git add only touched files ```

Error Handling

Comparison: Single-Agent vs Multi-Agent PDE