Known limits

We prefer to say it here rather than let you find out in the middle of a workshop.

Verified

• Missions: commands, outputs and cleanup ran in a real account on 06/11/2026, through the AWS CLI, with no resource left over.

• AI answers: one test per prompt, with links and citations checked.

• Tests of the CLI, the engine and the MCP server, the latter with the official client.

• Parity between core.mjs and core.py, output by output.

• Python tests of the lab, including the Strands SDK cycle with a simulated Bedrock.

• Five cases with real Bedrock, one sample per case.

• Two Docker images built and run.

Not yet verified

• Missions in CloudShell and in the console: the commands ran through the AWS CLI, not inside CloudShell, and the console clicks follow the documentation.

• Missions with beginners: no group has tested them yet. Times and difficulty are estimates.

• Free tier values: they change and depend on the account plan.

• AgentCore managed runtime: it was not deployed or rehearsed.

• Inside Kiro, Cursor and VS Code: the format follows the official documentation, with no test in the app.

• Reproduction by other people: the report form is open for that.

• Workshops with a real group: the times are estimates, not measurements.

Content limits

• The simulation is deterministic: it does not show how a model behaves.
• Passing case A4 does not prove resistance to prompt injection.
• The queue calculation uses constant rates. It is not a benchmark or a production configuration.
• check_citations checks citations, not the truthfulness of the answer.
• The limit of 4 calls in Bedrock mode is not a global spending cap.

Use

Do not use customer or production data. Data, incidents and metrics are all made up. Independent content, with no confirmed schedule at events and no official relationship with AWS.