When your company has 100 AI agents, who controls the permissions, the costs and the mistakes?

The catalog has to be searchable and linked to identity, recording the Owner, Sponsor, Version, Model, Knowledge, Tools, Environment, KPI, Cost and Dependencies. Statuses should include Draft, Testing, Approved, Suspended and Retired, with evidence of approval.

Set up an intake process for teams to propose use cases, and check the existing agents before allowing a new one to be built. Use shared templates and connectors to reduce duplication. Set an expiry date: if an agent goes unused or its sponsor doesn't confirm it, the system automatically reduces its permissions or shuts it down.

Never use shared accounts. Create a dedicated identity so you can trace which agent each action came from. Assign an accountable sponsor and grant permissions on a least privilege basis. Distinguish between agents acting on behalf of a user and autonomous agents that run with no person in the session.

Use time-bound access, approvals and conditional policies. An agent reads only the data its task requires and never gets admin rights just to make integration easier. Secrets must be kept in a vault and be rotatable. When a sponsor changes jobs, ownership has to be transferred or the agent suspended automatically.

The risk tier determines evaluation, monitoring, approval and how often access is reviewed. Summarizing a meeting shouldn't go through as much process as approving a loan, but every tier still needs an owner and a defined set of permitted data.

Logs link the request, plan, tool calls, data sources, policy decisions, human approvals and outcome. Use a trace ID to follow work across agents. Don't keep more confidential data than necessary, and set retention according to risk.

The dashboard shows success, exceptions, latency, cost, policy violations and outcomes. Raise an alert when an agent behaves in an unusual pattern, uses tools abnormally often or drifts in quality. Replay samples and test golden cases with every release.

Multi-step agents may call models and tools over and over. Set budgets per task, per agent, per team and per month. Use small models for classification and extraction, and expensive models only for reasoning that really needs them. Cache data and limit loops.

Define a playbook: Detect, Contain, Revoke, Rollback, Notify and Learn. Have a kill switch for each agent and tool, plus a global emergency mode. Run tabletop exercises, for example an agent sending out large amounts of wrong data, or leaked credentials.

Keep a manual fallback and minimum capacity for critical workflows. Back up configuration, prompts, policies and data lineage. A vendor outage must never leave the company unable to see the status of its work. Set RTO and RPO according to impact.

Use a federated model: a central platform and security team looks after identity, policy, observability and the catalog, while domain teams own the workflows, knowledge and outcomes. Set up an AI/Agent Council for standards and high-risk cases; it shouldn't be approving every daily task.

Review the portfolio quarterly and decide whether to scale, improve, merge or retire each agent. Measure register coverage, access reviews and incidents together with business value. Build a red team or quality team that tests important agents independently of the people who built them.