Harness Levers
Each project changes one lever. Keep the task stable, change the harness, then inspect whether behavior improves.
The Levers
| Lever | Question | Project |
|---|---|---|
| Trace | Can you explain what happened? | P01 |
| Model | Did the task need the expensive model? | P02 |
| Retrieval | Did extra search improve the answer? | P03 |
| Decomposition | Did smaller prompts beat one large prompt? | P04 |
| Memory | Did durable context reduce re-discovery? | P05 |
| Safety | Did the harness bound the blast radius? | P06 |
| Critic | Did evaluation stop false completion? | P07 |
| Workflow | Did model-authored orchestration reduce glue code? | P08 |
| Escalation | Did routing recover when cheap got stuck? | P09 |
| History | Did indexing past sessions beat scanning them? | P10 |
The Rule
Excessive harness features can lead to bloat. Use traces and metrics to ensure your harness is optimized for your work.
That is why the course teaches: predict, run, inspect, measure, keep.