Results from the lecture examples

Open the lecture examples to inspect the full prompts, actions and returned evidence. Only the current lecture selections are included.

• Workshop 1: compare how OpenHands, OpenCode and Pi inspect and check the same repair.
• Workshop 2: compare terminal retrieval, GitHub operation catalogs and a custom evidence-bundling tool.
• Workshop 3: both the no-memory and useful-memory runs pass all four final tests. The saved note changes the first check selected, not a demonstrated speedup. The stale-memory example is an earlier take with a different reuse-first workflow.
• Workshop 4: both models produce animated SVGs, and both outputs have motion flaws. One pair does not establish a general model ranking.
• Workshop 5: external feedback identifies a deliberately missing report record. The agent repairs the report and the finish-time check passes. This demonstrates external verification; it does not establish a benefit over a baseline or native goal persistence.
• Workshop 6: inspect the single-owner run alongside the lead and three workers. Both saved patches pass checks, but the team lead reaches a capped stop. That is not a completed team workflow.

Full selected histories, documented redactions, original model labels and task-specific evidence are included in the lecture download. Use these observations to design your own comparison rather than treating a selected example as a general performance claim.