Reproducible runs, improvable steps
A prompt lets the harness invent the route, so every run is a different run and every score a one-off. A blueprint fixes the steps, so you can study them one at a time and raise the score on purpose.
this route runs every time, in this order
| iter | what changed | score | delta |
|---|---|---|---|
| 01 | baseline | 0.62 | |
| 02 | retrieve → v2 | 0.71 | +0.09 |
| 03 | rank removed | 0.66 | −0.05 |
| 04 | draft → v2 | 0.86 | +0.15 |
That is what reproducibility buys: a system you can study a part at a time, and improve on purpose rather than by luck.
What surrounds a run →This is a blueprint
Which agents run and what each hands to the next. Download the folder and run it.
A blueprint pins the handoffs, loops, checkpoints, and deliberate absences that make a workflow reusable. The files stay plain enough to inspect before your harness runs them.
What a blueprint is →Every node is a card
Open one and it says what it does, the brief it is handed, which model runs it, what arrives, and what must never reach it.
Work through the build brief and emit the source it describes, adding nothing the brief does not ask for.
Each node pins an exact card version: its job, interface, tool reach, and prohibitions. Reuse the card in another graph.
Card format reference →Write your first blueprint
The tutorial takes you from a small task to a folder your agent can run, and the graph draws itself as you answer.
- 01 Design in conversation
- 02 Watch the graph draw
- 03 Enrich through MCP
- 04 Keep it on your account
In Claude Code or Codex. No account needed to start.