Engineering the Harness: six practical labs
Full slide deck (draft, October 19)
Choose how you want to participate. We recommend inspecting the saved conversations in your own Agent Canvas. You can follow the recorded decisions and tool results without waiting for a new run or paying for model calls.
| Option | What you do | When to choose it |
|---|---|---|
| 1. Read the HTML | Open the saved transcripts in your browser. | Quick access with no Canvas setup. |
| 2. Inspect in Agent Canvas (recommended) | Import the lecture histories, filter the conversation list, and expand tool calls and results. | Follow along with the instructor and investigate the traces yourself. |
| 3. Run the full exercises | Start fresh conversations and perform the tasks with your own model access. | You have time to configure the tools and compare new runs. |
Follow Get started to choose your path. Use saved-conversation directions for the recommended Canvas import, or Exercise prompts to run the work yourself.
| Lab | What you do |
|---|---|
| 1. Harness choice | Repair the same pagination bug in separate harnesses |
| 2. Tool design | Explain a PR with different GitHub tool surfaces |
| 3. Memory | Recover a prior customer agreement from different supplied notes |
| 4. Model routing | Compare counter and controlled security repairs; verify before escalating |
| 5. Goals and stopping | Inspect a staged report checkpoint, rejection and continuation |
| 6. Delegation | Compare one reservation owner with independent adversarial discovery |
You can run the exercises yourself or inspect saved traces from my agent runs. Use the worksheet to record your predictions and observations.
Make your predictions before reading the saved results.
The slides show recorded experiments. These labs let you explore the same ideas with your current models and tools; your action sequences and results can differ.
Afterward, use the optional long project or select a project from the full course. The talk's conceptual companion remains harness-engineering.