A while back I wrote about letting an AI agent do your job searching. That post ended with a quirk I couldn't explain. EventBridge Scheduler would invoke my agent, the agent would run fine, I'd get the email, and the invocation would land in the dead letter queue anyway. Every single time.
I figured I was either doing something wrong or hitting an AgentCore bug. Turns out it was neither.
The Actual Problem
EventBridge Scheduler universal targets (the feature that lets Scheduler call almost any AWS API action directly, no Lambda in between) are synchronous. Scheduler makes the API call and waits for a response before deciding whether the invocation succeeded. The docs describe this as at-least-once delivery, where at least one delivery succeeds with a response from the target. I glossed over that last part.
My schedule was calling the Bedrock AgentCore invokeAgentRuntime action, which blocks until the agent finishes. A search takes 30 to 75 seconds. Scheduler gives up around the 30 second mark and marks the invocation failed. The agent, which has no idea any of this happened, keeps going, finishes the search, and publishes to SNS. I get my email. The "failed" invocation goes to the DLQ. Here's what one of those messages looks like:
