slb

A two-person rule workflow requires a second reviewer to explicitly authorize potentially destructive AI agent commands before execution. Approvals are cryptographically bound to exact command content via SHA-256 hash, preventing tampering or substitution in multi-agent coding environments.

Command hash binding prevents unauthorized execution by attaching each approval to the exact command content using a SHA-256 hash. This ensures any modification to the command after approval invalidates the authorization, blocking substituted or tampered commands from running in production environments.

You can classify AI agent commands into CRITICAL, DANGEROUS, CAUTION, and SAFE risk tiers, each with tailored approval rules. This risk-tier classification allows differentiated governance, ensuring only commands meeting specific risk thresholds trigger mandatory two-person review workflows.

A notary daemon is used to verify that command approvals are valid before client-side execution. Commands run in the user’s environment with preserved credentials and context, while the daemon independently confirms that the required two-person review and hash binding are satisfied.

Auditing decisions and outcomes for AI-initiated commands is handled through persistent logs and state that record all approvals, decisions, and execution results. This creates an auditable trail showing patterns and governance outcomes across multi-agent coding workflows over time.

A two-person rule for AI command approval may be unnecessary for SAFE-tier commands that pose no risk to data, infrastructure, or production environments. Over-applying peer review to low-risk operations creates workflow friction without adding meaningful safety value to the multi-agent pipeline.