The Operator Path
Four levels. Measured by what you keep.
Most capability ladders measure what the machine produced. This one measures what you can still do when it is switched off — because that is the part that is actually yours.
The four levels
- 01
Generate
You can make things with a model.
You get useful output on demand — drafts, code, analysis — and you can tell good output from bad in a domain you already know.
- What the system carries
- Nothing yet. At this level the model is a tool you pick up, not a system that runs. That is the correct place to start.
- What you keep without it
- You can still do the work yourself, slower. Your judgment about quality is your own, not borrowed from the output.
- The receipt
- Work you shipped and would defend in a room with no laptop.
- Where people stall
- Output volume rises and nothing compounds. Every task starts from an empty prompt because nothing was ever written down.
- 02
Direct
You can aim it before it runs.
You state intent, constraints, and what would count as done — before the system answers. You write acceptance criteria you are willing to be held to.
- What the system carries
- Your standards become files. A skill captures how a job is done well; a rubric captures what proof it must return.
- What you keep without it
- You can articulate the standard from memory. If the file vanished you could rewrite it, because you decided it — you did not discover it in the output.
- The receipt
- A SKILL.md and an EVAL_RUBRIC.md that another person can run and get your quality.
- Where people stall
- You direct one thing at a time and become the bottleneck the moment two run at once.
- 03
Orchestrate
You can run several at once and keep them honest.
You decompose work across specialists, give each a scope narrow enough to be good at, and define where they stop and hand back to you.
- What the system carries
- A dispatcher routes work. Agents carry roles with written authority and stop conditions. Gates catch failures before you see them.
- What you keep without it
- You can explain why each boundary sits where it does. You would notice if a gate stopped firing, because you know what it was protecting.
- The receipt
- An AGENTS.md with real authority boundaries, and run receipts showing gates that caught something.
- Where people stall
- The fleet runs, but nothing it learns survives the session. Quality depends on who happened to be steering.
- 04
Steer
You can govern the system over time.
You shape what the system remembers, how it is evaluated, and how it corrects — including deciding what it must never do without you.
- What the system carries
- Memory persists with a privacy boundary you set. Evals run against a moving target. The system escalates rather than guesses.
- What you keep without it
- This is the hard one. You can still make the consequential call yourself, and defend it against a system that disagrees with you.
- The receipt
- A MEMORY_MAP.md, an eval history across dates, and at least one decision where you overruled the system and were right.
- Where people stall
- None yet that we have measured honestly. Anyone claiming a fifth level is selling something.
The honest part
We do not yet know if this works.
The ordering above is our working model, drawn from operating a fleet daily. The claim we are not making is that using our system develops the operator — nobody has measured that, including us, and we have the strongest possible conflict of interest in the answer.
It is an open question on our research agenda: does sustained agent assistance degrade your ability to independently defend the conclusions the system produced? A system that makes you faster while quietly making you weaker is a bad trade, and it would look exactly like success from the outside. That is why every level here is written as something you can still do with the system switched off — and why we would rather publish the measurement than the promise.