# What is Agent Constitution?

Also called operating principles, agent charter.

An agent constitution is a written set of principles that govern an agent's behavior, stating what it must always do, never do, and how to resolve conflicts between goals. In deployed systems it usually appears as durable instruction text and enforcement rules, distinct from the research method of training a model against a written set of principles.

The word carries two meanings that are worth separating. In published research, constitutional AI describes a training method in which a model critiques and revises its own outputs against explicit written principles, with the resulting preferences used for training. In everyday agent engineering, a constitution more often means a runtime document of standing rules that the agent is instructed to follow.

A runtime constitution differs from ordinary task instructions in scope and lifetime. Task instructions change with each request, while constitutional rules apply to every request and are meant to hold even when a specific task suggests otherwise. Ordering matters most where rules conflict, so a usable constitution states which principle wins rather than listing rules as equals.

Instruction text alone does not enforce anything. A rule saying an agent never sends external messages without approval is only as strong as the model's compliance, which adversarial or confusing input can erode. Rules that carry real consequences should also exist as code, meaning a permission check, an approval gate, or a validator that inspects the action before it executes.

Constitutions decay when they grow. A short list of genuinely non negotiable rules is read and followed, while a long document accumulates situational guidance that dilutes the important lines and consumes context on every call. Situational rules belong closer to the task, leaving the constitution to hold only what must never vary.

## Key points

- Standing principles that apply across all tasks
- Distinct from constitutional AI as a training method
- Must state precedence where rules conflict
- Text alone is not enforcement; back critical rules with code
- Short lists hold; long ones dilute and consume context

## In practice

A finance operations agent carries five standing rules, including one stating that it never issues a refund without an approval record and one stating that when a rule conflicts with a user request, the rule wins and the conflict is reported. The refund rule is also enforced by an approval gate in the payment tool, so a persuasive instruction in a customer email cannot bypass it.

## Related terms

- [System Prompt](/en/glossary/system-prompt)
- [Agent Role](/en/glossary/agent-role)
- [Guard Condition](/en/glossary/guard-condition)
- [Human in the Loop](/en/glossary/human-in-the-loop)
- [Escalation Policy](/en/glossary/escalation-policy)

[Back to the AI Glossary](/en/glossary)