The Levels of AI Orchestration
L0 to L7 — what the human actually does at each rung, how I work there, and where I honestly sit. Scroll to climb.
This ladder stands on prior art. Those frameworks measure a system's autonomy. This one measures a different axis — how much of the loop the human still occupies, and what they built to hand the rest away.
- Taxonomy & Definitions for Driving Automation Systems · SAE International, 2021 · J3016_202104
Six levels of driving automation (L0–L5). The shape every 'levels' ladder borrows — by analogy, not inheritance.
- Levels of AGI for Operationalizing Progress on the Path to AGI · Morris et al., Google DeepMind, 2023 · arXiv:2311.02462
Defines a six-rung Levels of Autonomy (Tool → … → Agent), and itself cites SAE as the precedent for leveling.
- Levels of Autonomy for AI Agents · Feng, McDonald & Zhang, 2025 · arXiv:2506.12469
Five levels keyed to the user's role (operator → observer). The closest kin — but scoped to one agent, not a fleet.
Chat / copy-paste
Open a model in a browser tab; paste code in, paste the answer back.
I kept a tab open and moved text by hand. The editor had no idea the model existed. I climbed past this the moment the model could see my files instead of my clipboard.
Autocomplete & approve-each-step
Inline suggestions you accept or reject; an agent that asks before every action.
The model shortened my keystrokes but I still drove every one of them. Every action waited on my approval. Useful — and not yet orchestration.
One autonomous agent, supervised
Hand an agent a whole task; let it edit, run, and iterate while you watch.
I gave an agent a task and let it run — intervening on drift, not approving each step. The unit of my attention became the task, not the keystroke.
A few parallel agents
Two or three sessions at once; context-switch between them, reviewing and unblocking.
My scarce resource stopped being typing and became attention. I ran a few sessions in parallel and moved between them — the first time the work outran one cursor.
Multi-machine fleet, away-from-keyboard
Many agents across machines, controllable remotely. You route, prioritize, resolve conflicts.
Agents across a Linux box and a remote Mac, driven over Tailscale. Air-traffic control, not typing — and the work keeps moving when I step away from the desk.
Platform-builder
Engineer the substrate the fleet runs on — gates, coordination, telemetry, independent review.
This is where I actually sit. I build the substrate the fleet runs on: hooks that gate on real exit codes, cross-machine file locks, telemetry that records everything, and an independent reviewer drawn from a different model family so its blind spots aren't mine. The substrate is the work; the agents are what runs on it.
- 38.9k tool calls
- 0.04% tool-call error rate
- 476 autonomous subagents
- different model family reviewer
Autonomy-platform owner
The control plane becomes something other people can trust and operate.
Locked — toggle “preview the gap” to read it.
Recursive / self-directing
The system improves its own orchestration; humans set objectives and constraints.
Locked — toggle “preview the gap” to read it.