Human in or out of the loop: autonomy levels & guardrails

Yesterday you closed the loop. Today you decide who drives it — and build the rails that let you hand over the wheel without crashing.

For eighteen days you built a machine. The genome that labels every creative at birth (Day 4). The matrix that turns one idea into fifty tagged assets (Week 2). The AI production line that ships them (Week 3). A clean test structure and a fitness function that defines "winner" by the deepest money event (Day 16). The dissection that reads winners down to elements (Day 17). And yesterday, the recursion: insights flow into the Creative Memory and brief the next batch, so quality compounds batch over batch (Day 18).

Now the operational question that decides whether this is a job or a system: which steps does a human still touch, and which run themselves? Get this wrong in one direction and you never scale — you're a very organised manual labourer. Get it wrong in the other and you scale a mistake at the speed of compute. Today is about finding the right rung, and the rails that let you climb safely.

1The four autonomy levels

Think of the loop as a car and yourself as the driver. Some days you want both hands on the wheel; some days the system can hold the lane while you watch the road. The dial has four rungs. At every rung the loop runs — what changes is how much of it the human does versus the system.

◀ More control, less scale  ·  The dial: how much of the loop the human runs vs. delegates  ·  More scale, more earned trust ▶

Read the dial top to bottom and you'll notice the human never leaves — they just move up the value chain. At L0 the human is a builder. At L3 the human is a policy-setter and auditor. The mechanical work (generating variants, applying genome tags, shifting €5 from a laggard to a leader) drains downward into the system at every rung. The judgment work (is this on-brand? is this claim true? is this winner real or noise?) stays human all the way up. That is the whole design principle: delegate the mechanical, retain the judgment.

Crucially, the system's two loops don't change with autonomy. Meta's inner loop — Advantage+ Creative and the auction allocating budget among the creatives you gave one ad set (Day 16) — runs automatically regardless of your rung. What the dial governs is your outer loop: who decides what to make next. L0 means a human writes every brief; L3 means the Creative Memory writes it and a human audits. The recursion from Day 18 is identical at every level — only the hands on it differ.

2The guardrails that let you climb

You cannot climb the dial on optimism. You climb it on guardrails — the hard rails that make delegation safe. The logic is simple: the more of the loop you let run unattended, the more constraints you must pre-commit so the unattended parts can't wander somewhere expensive or off-brand. Guardrails are what convert "I hope the system behaves" into "the system cannot do the thing I'm afraid of." Here is the standing set.

Notice the pattern: most rails are system-held (the machine enforces them on itself), but the two that touch taste, ethics, and truth — the compliance gate and the veto — are human-held, and they stay human at every rung including L3. A guardrail isn't a suggestion the system can override when convinced; it's a wall. The whole reason you can sleep while an L3 loop runs is that the worst outcome is already bounded before you went to bed.

3How to choose your rung — and how to move up

Your level is not a personality; it's a function of three inputs: risk tolerance (how expensive and how reversible is a bad batch?), volume (how many batches per week — is manual even physically possible?), and maturity (how rich is your Creative Memory, and how good is the system's track record?). A new account with a thin Memory and a regulated product belongs at L0–L1. A six-month-old engine with a fat Proven Elements Library, a populated Graveyard, and three batches of clean compounding has earned L2, and is a candidate for bounded L3 on its lowest-risk cells.

Here's a concrete read of how you'd actually move up. Suppose your engine ships 30 tagged variants a week across 5 concepts. At L0/L1, a human reviews all 30 plus does the buy and the read — call it ~6 hours of human attention per batch. The Creative Memory is young: maybe 4 proven elements scored, win-rate noisy at ~12%.

Three months on, the Memory has 19 proven elements segmented by persona, a Graveyard of 30+ dead ideas you'll never re-test, and win-rate has climbed to ~24% with CPA trending down batch over batch (exactly the compounding you proved in Day 18). Now the system's brief-from-Memory is demonstrably as good as the human's, and the dissection it drafts matches the human's read 9 times in 10. You move to L2: the system launches and allocates inside a €150/day batch cap, auto-kills any variant 30% below the hook-rate average after 48 hours, and you only show up at two checkpoints — crowning winners and the brand gate. Human attention per batch drops to ~90 minutes, and throughput doubles because nothing waits on your inbox.

That is the move: you earn each rung with evidence. The promotion criterion is never "I'm busy" — it's "the system's autonomous decisions have matched my judgment, on this kind of cell, often enough that I trust the guardrails to catch the rest." You ratchet up one rung, on one low-risk cell type, watch it, then widen. You never wake up one morning and flip the whole engine to L3 because a dashboard told you your Opportunity Score would rise.