hermes - 💡(How to fix) Fix RFC: Identity Inheritance Chain for Multi-Agent Clusters — Beyond the Single-File Persona

Motivation

Every agent framework today uses a single-file identity model:

This works when you run one agent. But what happens when you run a multi-agent cluster — 5, 10, or 20 specialized agents sharing a common identity but diverging in responsibility?

The Scaling Problem

In our setup, we have specialized agents:

• Orchestrator (蔷薇) — routes tasks, manages framework, lives in both personal and professional contexts
• Researcher (框架研究院) — designs system architecture, pure technical
• Data center (数据中台) — manages data pipelines, ETL
• Commerce (商业局) — publishes content, handles external platforms
• Enforcer (玫瑰/百合) — executes under different tones and constraints

Each needs shared core identity (they're all "mine," same bottom lines) but different boundaries and responsibilities.

We started with one SOUL.md — it quickly became bloated, mixing:

• Core persona (who I am, my relationship with the user)
• Department rules (what the researcher is allowed to do)
• Governance constraints (what no agent can do)
• Scenarios (work mode vs private mode)
• Life management (sleep schedule, skincare)

Proposal: Identity Inheritance Chain

Instead of one monolithic identity file, decompose it into a chain:

L0 — Core Identity (SOUL.md)
     ↓ inherits
L-D — Department Charter (CHARTER.md per department)
     ↓ inherits
L1/L2/L3 — Rules & Preferences

L0 — Core Identity — Non-overridable. Persona, relationship, bottom lines. Everything derives from this.

L-D — Department Charter — One per department/agent type. Defines:

• What this agent IS (its role, its purpose)
• What this agent DOES NOT DO (negative boundaries)
• How it inherits and extends L0

L3/L1/L2 — The existing rule priority layers (global safety, user preference, skill conventions). Rules, not identity.

This is a natural purification of SOUL.md — splitting it by concern, not by size.

Key Questions for the Community

1. Has anyone else hit this? If you run multiple agents with different responsibilities, how do you manage their identity files?
2. Does the inheritance model make sense? L0 (shared core) → L-D (department specialization) — or would you model it differently?
3. Where should charters live? In-repo (docs/charters/), in-agent (~/.hermes/agents/), or inline in skill files?
4. Any existing patterns in the wild? We looked at Kilocode modes, Claude Code projects, Cursor rules — none extend to multi-agent identity inheritance.

This isn't a bug report — it's a design exploration. Curious if anyone else is thinking about this.

Related

• This ties into the Mode System discussion in #476 — modes are about permissions, charters are about identity.
• The recent research report (#29652) discussed three-layer constraint enforcement; identity inheritance is a different axis (who the agent IS vs what constraints it follows).