Product & Strategy
MDX
A Practical Governance Framework for AI Agents
Assign ownership and controls based on capabilities, data access, action impact, and evidence of reliability.
Editorial review: clarity, operational relevance, safety boundaries, and source quality.
Inventory capabilities
Record each agent's purpose, users, models, tools, data, permissions, and deployment owner. Governance starts with knowing what can act and where.
Maintain a living inventory for every deployed agent: purpose, owner, users, models, tools, data classes, credentials, environments, autonomy level, vendors, evaluation status, and incident contact. Include internal experiments once they touch real organizational data or systems.
Tier by risk
Classify use cases using consequence, reversibility, data sensitivity, and degree of autonomy. Higher tiers require stronger testing, approvals, monitoring, and review cadence.
Risk tiering should combine impact, reversibility, scale, data sensitivity, external exposure, and the agent's freedom to choose actions. Map each tier to minimum controls such as independent evaluation, security review, approval boundaries, trace retention, launch gates, and reassessment frequency.
Treat evidence as ongoing
Approval at launch is not permanent proof of safety. Reassess when models, prompts, tools, policies, or usage patterns change, and learn from near misses as well as incidents.
Material changes include new models, prompts, tools, permissions, data sources, regions, user populations, or task scope. Tie deployment records to evaluation evidence and approvals so governance follows the running system. Review incidents and near misses across teams to update common controls.
Practical example
Governance by capability
A document summarizer using public data is low tier and requires basic quality tests. Adding confidential contracts raises data controls; adding the ability to send notices raises action impact and requires approval, security review, and stronger monitoring. The product name is unchanged, but governance changes because its capabilities and downside changed.
Field checklist
Apply it in practice
- Inventory owners, tools, data, permissions, and deployment versions.
- Map risk tiers to mandatory technical and review controls.
- Require evidence before launch and material changes.
- Share incident lessons and retire unused authority.
Decision framework
Questions to answer before you build
Governance connects deployed capabilities to named owners, risk-based controls, current evaluation evidence, and a repeatable process for material changes and incidents.
Is every active capability inventoried?
Record purpose, owner, users, models, tools, data classes, credentials, environments, autonomy, vendors, and incident contacts.
What determines the risk tier?
Combine consequence, reversibility, scale, sensitive data, external exposure, and freedom to choose actions.
What triggers reassessment?
Treat new models, prompts, tools, permissions, data, regions, user groups, and task scope as material changes tied to fresh evidence.
Common failure signals
Watch for these warning signs
- Governing product names rather than the capabilities they currently possess.
- Treating launch approval as permanent evidence of safe operation.
- Creating review paperwork without technical enforcement or incident feedback.
Selected primary references
Continue with the source material
These sources inform the wider editorial perspective for this topic. They are not presented as line-by-line citations for every statement.