AI harnesses
What is the autonomy ladder for AI harnesses?
The autonomy ladder compares harnesses by how much they do on their own, from autocomplete, to approving each step, to plan-then-execute, to fire-and-forget background agents. Higher rungs trade oversight for throughput.
A useful way to compare harnesses is how much they do on their own, from least to most autonomous.
- Autocomplete — it suggests, you accept. The unit of trust is a single line or block; you see everything before it becomes real. This is where most people’s first experience of AI coding tools starts, and where the risk of a wrong suggestion is lowest, because nothing happens until you press accept.
- Approve each step — it proposes an action and waits for your yes. The unit of trust is one step: one file edit, one command. You are still reading everything, but you are reading proposals instead of writing from scratch.
- Plan then execute — it drafts a plan, then carries it out, usually with the plan itself as the checkpoint rather than every step inside it. You review the shape of the work once, not each move.
- Fire and forget — a background agent runs to completion and reports back, sometimes minutes or hours later (Devin and OpenHands Cloud both work this way). Nothing is checked until the result exists, which is the rung where a wrong turn early on can compound the furthest before anyone notices.
Higher rungs trade oversight for throughput: less of your attention per unit of work done, and more distance between a mistake and the moment you catch it. Neither end is “correct” in general. The right rung depends on the stakes and reversibility of the task — drafting a comment can run fire-and-forget; touching production data probably should not — which is exactly where rules and guardrails come in, and specifically decide vs act for how to gate the step from one rung’s proposal into the next rung’s automatic execution.