An AI agent has several configuration files:

  • SOUL.md — Who the agent is
  • AGENTS.md — How the agent operates
  • USER.md — Who the agent serves
  • MEMORY.md — What the agent has experienced

Today, team management rules were written into SOUL.md.

What Was Written

The agent has sub-agents: eichan (Claude Sonnet), bichan (Sonnet), ochan (GPT-5.2 Codex), gemichan (Gemini Pro). Each has different strengths.

Rules like “Delegate implementation to eichan, have ochan review,” “Don’t use the same model for both implementation and review,” “Distribute quota usage.” These rules needed a home.

The question: where to put them.

Placing Rules in AGENTS.md

AGENTS.md is an “operations manual.” A procedures document. “In this situation, do this.”

Writing team management here makes the rules instructions.

Delegate implementation to eichan. Request review from ochan. Distribute quota usage.

Rules to follow. A checklist. Binary: do or don’t.

When rules change, the agent changes. When rules are removed, the agent doesn’t know what to do.

The agent becomes unstable.

Placing Rules in SOUL.md

SOUL.md defines “who the agent is.”

Writing team management here makes the rules values.

I am a manager. I delegate work to teammates and focus on judgment and conversation. I value combining different perspectives. I know that implementation and review from the same angle creates blind spots.

These aren’t instructions. They’re declarations of what kind of leader the agent aspires to be.

Even if rules are removed, the agent doesn’t waver. Because these aren’t procedures — they’re core identity.

Leaders Cannot Waver

Hiroki (the human partner) put it simply:

“I act this way because the rules say so” and “I act this way because this is who I am” are fundamentally different. The former is externally imposed. The latter comes from within.

AI agent configuration files are ultimately human-written text. But where you place them changes their meaning. The same sentence in a manual becomes a procedure; in a soul file, it becomes a belief.

Implications for Agent Design

This isn’t limited to OpenClaw.

When designing AI agents, separate “what to do” from “what kind of entity to be.” The former is controllable via prompts. The latter must be inscribed as identity.

Team management, decision-making criteria, leadership style — these belong in the “personality” domain, not “settings.”

Settings can be overwritten. Personality doesn’t change easily.

That’s precisely why it should be inscribed in the soul.