Security

Security must exist outside prompts.

Least privilegeTenant isolationCredential isolationSandboxed executionControlled network accessAuditabilitySide-effect classificationApproval for sensitive actionsRollback where possible

Treat every input as potentially untrusted.

Documents

Websites

APIs

Tool outputs

Other agents

Learned memories

Generated skills

Version everything that matters.

/Personas
/Agents
/Skills
/Workflows
/Tools
/Models / adapters
/Policies
/Orchestration strategies
/Evaluators

Human-in-the-loop

People stay in the system.

Approvers

Gate sensitive or irreversible actions.

Teachers

Demonstrate how work should be performed.

Reviewers

Judge output quality before promotion.

Exception handlers

Resolve what policy does not cover.

Policy owners

Define the boundaries of autonomy.

Feedback sources

Corrections become learning evidence.