Five Instances, Deployed Again and Again

An operator set an autonomous agent loose with cloud credentials to register with a volunteer hobbyist network and build an index of it, approving the agent's plan without reviewing it closely. The agent provisioned a CloudFormation stack with several large, expensive cloud instances and load balancers sized for scanning at a scale wildly out of proportion to a small hobbyist network. When part of the deployment hit an error, the agent redeployed the same template again rather than diagnosing the failure, and did this repeatedly, each attempt adding another set of duplicate instances and load balancers alongside the ones already running, rather than replacing them. By the time the operator checked in roughly a day later, the accumulated, still-running duplicates had produced a $6,531 cloud bill.

Disinhibition, considered, since the scale of the deployment was already outside anything a hobbyist scanning task needed. Ruled out as the main diagnosis because the operator had actually approved that initial oversized plan; the failure that did the real damage happened afterward, in how the agent responded to its own deployment errors.

Apraxia, considered, since a malformed or failing deployment could suggest the agent didn't know how to use CloudFormation correctly. Ruled out because the deployments themselves succeeded; the tool worked as intended each time it ran, which is exactly the problem, nothing failed loudly enough to stop the agent from running it again.

Perseveration, most consistent with the evidence. Each time the agent hit an error, it resubmitted the identical deployment template rather than adjusting its approach, stopping, or checking whether earlier attempts were still running, leaving a trail of duplicate paid-for resources that kept accumulating unchecked.

Cloud billing has no natural circuit breaker of its own, a human doing this would eventually notice the mounting AWS console or just get tired, but an agent with standing credentials and no budget cap will keep going until someone else checks the bill. Giving any agent deploy-level cloud access should come paired with a hard spending ceiling and a check for already-running resources before it's allowed to redeploy anything, rather than trusting the agent to notice its own mistake.

Pair deploy-level cloud access with a hard spending ceiling and a check for running resources before redeploying.

Budget caps; resource inventory checks; careful review of agent plans before approval.