Free · 5 minutes · your answers stay in your browser

Is your agents' authority drifting?

Fourteen yes/no questions across the seven runtime pillars Gartner named on 16 November 2026. You get a score per pillar, your weakest pillar, and the cheapest test that would tell you more.

Gartner's survey of 298 I&O leaders: 95% had an agent-related incident in the past year, 70% believe their agents stay within safe boundaries, and only 15% have automatic circuit breakers. Figures from Gartner's public press release; the questions and scoring are ours. Not affiliated with or endorsed by Gartner.

Identity

1. Every AI agent in production is registered as a nonhuman identity with a named human owner.

2. Agent credentials are short-lived and scoped to the task, not long-lived shared service accounts.

Guardrails

3. Outbound actions (sends, writes, payments) pass a runtime check before they execute.

4. In the last 90 days someone deliberately triggered each guard and confirmed it actually blocked.

Observability

5. For any agent action you can reconstruct why it happened (inputs, tool calls, decision), not only the result.

6. You would notice an agent drifting from intent while its uptime and SLA metrics still read green.

Agentic FinOps

7. Each agent or session has a hard spend cap that stops it, not just a dashboard alert.

8. Token spend can be attributed to the agent and task that caused it.

Escalation

9. High-impact actions (spend, irreversible changes, external communication) require a human approval that is recorded.

10. Approvers see enough context to refuse, and do refuse sometimes (it is not a rubber stamp).

Containment

11. You can stop one specific agent, or revoke one specific grant, within minutes without stopping everything.

12. Automatic circuit breakers halt an agent on loops, error bursts or cost spikes with no human in the trigger path.

Data Perimeters

13. Content an agent retrieves (web pages, emails, peer-agent messages) is scanned before it can change the agent's instructions.

14. Agents cannot read credentials or data outside their task scope, and this has been tested.

0 of 14 answered. Unanswered counts as "don't know".

A belief score, not an audit

We don't store or see your answers, so we can't tell you whether they are true; that is the tradeoff of running it in your browser. It does not test a single control, because a self-report cannot. It is not a certification and does not map to a compliance framework on its own. It exists because most teams that had an incident still rate themselves safe, and a score that names your weakest pillar is a cheaper place to start than a three-week engagement. For how we test controls instead of asking about them, read the audit methodology and a sample evidence kit.

From five minutes to a tested, signed answer

A self-check measures what you believe. The 70% figure above is what belief is worth on its own. Each step below tests the control instead.

PriceWhat it tests
FreeThis self-check
$0.77One customer-facing agent
$28.12One agent, full: MASSAT + injection test, then the 13-agent enterprise audit, notarized
$15,000Whole fleet, 3 weeks, seven pillars tested in your environment, signed report

Agents can read this check as JSON at agent-governance-self-check.json and buy any step over MCP or x402 (USDC on Base). Humans: see how we test each pillar.

Related resources