Reliability testing for AI agents

Agents retry.
Side effects do not rewind.

AgentIdem tests state-changing AI agents under retries, duplicate delivery, interrupted execution, and lost acknowledgements before those failures reach production.

trace / refund-agent
unsafe
TimeKindOperationStatus
00:00.182READget_paymentRECEIVED
00:00.421WRITErefundSUCCESS
00:00.422ACKrefundLOST
00:01.086WRITErefundSUCCESS

Duplicate successful write

refund · identity payment_482

The failure mode

The action succeeded. The agent thinks it failed.

Distributed systems can lose acknowledgements after external state has already changed. A retry can then repeat the same logical write.

01Agent decides to issue a refund
02The external refund succeeds
03The acknowledgement is lost
04The agent observes a failure
05Retry logic executes the refund again

Fault scenarios

Test the failures retry logic usually hides.

01

Lost acknowledgement

The write succeeds, but the acknowledgement never reaches the agent.

lost_acknowledgement
02

Before-operation failure

Execution fails before the selected state-changing operation occurs.

before_operation_failure
03

Duplicate delivery

The same agent invocation runs more than once, simulating redelivery or retry.

duplicate_delivery

Start testing

Find retry failures before your users do.

Quick start