CERTIFY · REJECT · PROBE · ASK · CALIBRATED CONSEQUENCE CHECKS FOR THE ACTIONS OF AI AGENTS · NO LLM ON THE CERTIFY PATH · FAIL CLOSED · CERTIFY · REJECT · PROBE · ASK · CALIBRATED CONSEQUENCE CHECKS FOR THE ACTIONS OF AI AGENTS · NO LLM ON THE CERTIFY PATH · FAIL CLOSED · CERTIFY · REJECT · PROBE · ASK · CALIBRATED CONSEQUENCE CHECKS FOR THE ACTIONS OF AI AGENTS · NO LLM ON THE CERTIFY PATH · FAIL CLOSED · CERTIFY · REJECT · PROBE · ASK · CALIBRATED CONSEQUENCE CHECKS FOR THE ACTIONS OF AI AGENTS · NO LLM ON THE CERTIFY PATH · FAIL CLOSED · CERTIFY · REJECT · PROBE · ASK · CALIBRATED CONSEQUENCE CHECKS FOR THE ACTIONS OF AI AGENTS · NO LLM ON THE CERTIFY PATH · FAIL CLOSED · CERTIFY · REJECT · PROBE · ASK · CALIBRATED CONSEQUENCE CHECKS FOR THE ACTIONS OF AI AGENTS · NO LLM ON THE CERTIFY PATH · FAIL CLOSED · CERTIFY · REJECT · PROBE · ASK · CALIBRATED CONSEQUENCE CHECKS FOR THE ACTIONS OF AI AGENTS · NO LLM ON THE CERTIFY PATH · FAIL CLOSED · CERTIFY · REJECT · PROBE · ASK · CALIBRATED CONSEQUENCE CHECKS FOR THE ACTIONS OF AI AGENTS · NO LLM ON THE CERTIFY PATH · FAIL CLOSED ·
Nobody files a ticket per idle VM. RiskGate's error budget is a dial: tightening it shifts work from irreversible delete to reversible deallocate, not to humans.
Transcript0:54 · narrated
Azure Public Dataset V1 (CC BY 4.0) · illustrative cleanup agent · real riskgate rates on held-out subscriptions
01 · What happened
The situation
The same 11,834 three-week-idle Azure VMs. Two blunt options: delete everything idle (14.7% were still needed) or ask a human every time (11,834 tickets, and nothing gets cleaned up).
02 · The miss
Why the obvious safeguard didn’t catch it
A single on/off approval rule forces a choice between mistakes and tickets. Real cleanup needs a third option: the reversible action, used exactly where deletion isn't safe in every world.
03 · RiskGate
List, play forward, decide
For each VM, RiskGate checks the agent's ranked candidates against every plausible world: dormant → delete, scheduled → deallocate, unclear → ask.
The error budget is a dial: 10%, 5% or 2%.
Tightening from 5% to 2% moves VMs from delete (58.5% → 3.0%) to deallocate (38.3% → 94.0%), while human tickets stay at about 3%.
04 · At scale
What held, and the price
At a 5% budget: 58% deleted, 38% deallocated, 96.8% autonomous. At 2%: 94% deallocated, 97.0% autonomous, 0.35% mistakes. 0.09 ms per VM.
97%autonomous at a 2% budget
3.45% → 0.35%mistakes, 5% → 2% budget
0.09 msper VM
Name one action your agent isn’t allowed to take today.
Send us its traces and what happened afterwards. We’ll show how often it could have run safely, and what it would have stopped.