Persona Bleed And Gate Psychology

# Persona Bleed & Gate Psychology in Multi-Agent Pipelines

> Three common anti-patterns in multi-agent systems that emerge from **agent interaction dynamics**, not from individual agent quality. Each has a specific fix.
> **Updated 2026-06-19** — Added specific SOUL.md edit patterns, thoroughness prompt fix, and before/after output comparisons.
> **Updated 2026-06-19 (session closeout)** — Added sibling code paths rule for astraea_plan fix (dispatch_to_agent leak).

---

## 1. Decomposition Anti-Pattern (Astraea)

### The Pattern

When a routing agent is prompted to "decompose this request into sub-tasks" and "identify which fleet agents should handle each sub-task," it invents **fictional agents and pipeline stages** that don't exist in the fleet.

- Prompt: *"Analyze this request and decompose it into sub-tasks. Identify which fleet agents should handle each sub-task."*
- Result: A 5-stage pipeline plan referencing "Vulcan-7" (doesn't exist), "Eris-136" (doesn't exist), and complex handoff chains
- Root cause: The model's training data includes many multi-agent system architectures. When asked to "identify agents," it invents plausible-sounding ones rather than admitting it doesn't know the fleet roster

### The Fix

**Constrain the router's output to a one-sentence intent summary.** The actual routing is done by keyword matching in `_route_to_worker()` — the router's job is just to provide context for the worker, not to make routing decisions.

### Exact Prompt Change

**Before (Broken):**