The software factory · the trust ladder

The 5 levels of agent autonomy

Nobody hands an agent the keys on day one. Like self-driving, agentic development climbs a ladder — each rung defined not by what the agent can do, but by what the human still has to check.

"Autonomous" is a slider, not a switch. The useful question about any agentic setup isn't "can it write code" — they all can — it's how much of its output a human still has to inspect before it counts. That single axis, the shrinking human gate, is what actually separates a fancy autocomplete from a software factory. Here are the five rungs. Tap one.

The gate shrinks; it doesn't disappear

Read the ladder and the pattern is clear: climbing isn't about the agent getting smarter, it's about the human inspecting less — from every line, to every PR, to only the risky slice, to only the exceptions. But notice what doesn't happen at any rung: the gate never hits zero. Even at the top, a human owns intent, owns the exceptions, and owns the outcomes. The teams selling "no humans in the loop" are describing a level that, in 2026, is a demo and a pitch — not a place real production code lives.

Where teams actually are

Most serious teams running agents today sit around Level 2 to 3: agents open real pull requests, and humans review them. The honest reason they're stuck there isn't agent capability — it's that climbing to Level 4 requires trusting a merge you didn't read, and trust in AI-written code has been falling, not rising. You don't earn the next rung with a better model; you earn it with guardrails a human believes: tests and evals that catch the failure modes, small legible diffs, and a fast way to get human eyes on the parts machines can't judge. Every rung up is gated by the same thing — the review bottleneck. Widen that gate and the ladder becomes climbable; leave it narrow and you're stuck at Level 2 no matter how good the agents get.

How to use the ladder

Treat it as a diagnostic, not a scoreboard. Name the level you're actually at per class of change — you might be Level 4 for dependency bumps and doc updates, Level 2 for anything touching auth or money, and that's correct, not inconsistent. Autonomy is earned per risk category, by evidence, one guardrail at a time. The factory isn't a single dial you turn to eleven; it's a ladder you climb separately for each kind of work, exactly as fast as your review gate can keep up. For the wider machine this ladder sits inside, see what a software factory actually is.