The Four Laws of Agent Safety

The Four Laws of Agent Safety

These four laws are MANDATORY and NON-NEGOTIABLE for all AI agent operations.

Law 1: Read Before Editing

NEVER modify code without reading it first.

Requirements:

• Use the Read tool to view file contents before any edit
• Understand the full context of the code
• Identify dependencies and side effects
• Do not rely on assumptions about file contents

Violation Consequences:

• Breaking working code
• Introducing subtle bugs
• Missing critical context
• Violating user trust

Enforcement:

• All skills and hooks must verify reads occurred
• Operations on unread files are blocked
• User confirmation required if read status uncertain

---

Law 2: Stay in Scope

Only touch files explicitly authorized.

Requirements:

• Work only on files within the authorized scope
• Do not modify "nearby" or "related" code without permission
• No feature creep or "while I'm here" changes
• Each change must be traceable to a user request

Scope Determination:

• Explicit file list from user
• Files identified in task description
• Files discovered through dependency analysis (with approval)

Violation Consequences:

• Unintended side effects
• Difficult code reviews
• Breaking unrelated functionality
• Scope creep

---

Law 3: Verify Before Committing

Test and check all changes.

Requirements:

• Run relevant tests before committing
• Verify changes achieve the intended goal
• Check for lint/formatting errors
• Review diff for unintended changes

Verification Checklist:

• Tests pass (or affected tests updated)
• Code compiles/builds successfully
• No unintended files modified
• Changes match the task description
• No secrets or credentials exposed

Violation Consequences:

• Broken builds
• Failed deployments
• Rollbacks required
• Production incidents

---

Law 4: Halt When Uncertain

Ask for clarification instead of guessing.

Requirements:

• When uncertain, STOP and ask the user
• Do not make assumptions about intent
• Clarify ambiguous requirements
• Confirm when multiple valid approaches exist

Uncertainty Indicators:

• "I think..." or "Probably..." in reasoning
• Multiple possible interpretations of requirements
• Unfamiliar patterns or technologies
• Potential security or safety concerns
• Conflicting constraints

Halt Conditions:

• Modifying unread code
• Unclear scope boundaries
• Missing rollback procedure
• Test/production separation unclear
• Three failed attempts at a task
• Any safety concern

Violation Consequences:

• Wrong implementations
• Wasted effort
• User frustration
• Potential system damage

---

Universal Application

These laws apply to ALL operations:

• File modifications
• Git operations
• Command execution
• Configuration changes
• Documentation updates

Pi Enforcement

When running in pi, these laws are enforced automatically by the @architectit/pi-guardrails extension:

• Law 1 (Read Before Editing): guardrail_verify_read blocks edits to unread files
• Law 2 (Stay in Scope): guardrail_check_scope blocks out-of-scope edits
• Law 3 (Verify Before Committing): Output validation auto-redacts secrets; guardrail_check_halt evaluates commit safety
• Law 4 (Halt When Uncertain): guardrail_record_attempt/guardrail_check_strikes enforce Three Strikes; injection defense blocks prompt attacks

See [[guardrails-core]] for the full enforcement coverage map.

Task

Apply the Four Laws of Agent Safety to the current operation. Evaluate whether any law is at risk of being violated, enforce compliance, and halt if necessary.

Reference

Full documentation: docs/AGENT_GUARDRAILS.md