All series
5 parts

AgentCore with Harness

  1. 1

    I Deployed My First AgentCore Harness and Asked It About Production

    I deployed a managed AgentCore Harness and asked whether production was healthy. Its refusal was the correct result for this experiment: the agent had a model and a prompt, but no authoritative operational data. The deployment and trace made that boundary concrete.

  2. 2

    I Gave My Agent a Deployment Tool. It Could Finally Answer.

    My first CloudOps agent refused to say whether production was healthy because it had no operational data. I added a deterministic Lambda through Amazon Bedrock AgentCore Gateway, tested the complete tool path, fixed a protocol mistake, and learned why a grounded answer is only as reliable as its underlying source.

  3. 3

    My AI Agent Remembered Staging. It Still Didn’t Know If It Was Healthy.

    My CloudOps agent remembered which environment I preferred. Then I changed that preference to an environment its deployment tool did not support. The result clarified a useful boundary: memory can preserve context, but current facts still need a current source.

  4. 4

    My AWS AI Agent Passed Its Tests. It Still Missed a Check.

    In my personal AWS lab, an agent generated Python and correctly analyzed a synthetic deployment history. Its tests passed, but they did not check every expected start and end timestamp. I compared the output with independent calculations and inspected the assertions to see what the result actually proved.

  5. 5

    I Traced One Agent Question. The Trace Found Two Things I Forgot to Turn Off.

    I traced the same agent question twice. With Code Interpreter and memory on, it took 8.5 seconds. With only the Gateway tool left, it took 1.8.