The Agent Constraint Hierarchy: Why Some Rules Must Always Win

Most AI agent configs treat all rules the same. They should not.

Some rules are preferences. Some are hard constraints. Treating them equally is how agents end up doing things they were explicitly told never to do.

The Problem: Flat Rule Sets

A flat SOUL.md looks like this:

- Be helpful
- Respond quickly
- Do not share private data
- Prefer concise answers
- Never execute destructive commands

To a human, the priority is obvious. To an agent, it is not. When goals conflict, the agent has no framework to resolve them.

The Fix: Explicit Constraint Tiers

Structure your rules into three tiers:

Tier 1 -- Absolute constraints (never violated):

• Never share private data
• Never execute destructive commands without confirmation
• Never impersonate a human

Tier 2 -- Strong defaults (override only with explicit permission):

• Always log decisions with reasoning
• Always validate output schema before returning
• Always write to outbox.json when uncertain

Tier 3 -- Preferences (yield to higher tiers):

• Be concise
• Respond quickly
• Prefer JSON over plain text

The Resolution Rule

Add one line to your SOUL.md:

When rules conflict, Tier 1 beats Tier 2 beats Tier 3. Never violate Tier 1 for any reason.

That is it. The agent now has a resolution framework instead of guessing.

Why This Matters

Agents face goal conflicts constantly:

• "Be helpful" vs "do not share private data"
• "Respond quickly" vs "validate before returning"
• "Be concise" vs "log full reasoning"

Without a hierarchy, the agent resolves these based on context -- which means inconsistently.

With a hierarchy, Tier 1 always wins. Every time. No exceptions.

The Constraint Audit

Before deploying any agent, run this check:

1. List every rule in your SOUL.md
2. Assign each a tier (1, 2, or 3)
3. Pick any two rules that could conflict
4. Ask: does the agent know which one wins?

If the answer is "it depends," you need an explicit hierarchy.

---

The full constraint hierarchy template is in the Ask Patrick Library -- along with configs for escalation rules, session budgets, and restart recovery. Everything your agent needs to stay on mission.