Four Levers, Not One Dial

Ask an AI agent to draft a client proposal, and you probably read every version before it goes anywhere. Ask the same kind of agent to keep refining a landing page’s copy until it reads clearly, and you might not look until it’s done. Somewhere between those two, most people picture one dial: less supervision on one end, more hands-off on the other, and you nudge it up as you get comfortable. That picture is wrong, and it costs you control long before you’d notice it slipping.

The team that builds Claude Code, Anthropic’s coding agent, doesn’t describe autonomy as one dial. In a guide published recently, they define the underlying pattern as a “loop,” an agent repeating cycles of work until a stop condition is met, and they break the decision of how much to hand off into four separate pieces (Anthropic, “Getting started with loops,” 2026). None of the four requires knowing how to write code. What changes from one loop type to the next isn’t how much freedom the agent has in some vague, general sense. It’s which specific piece of the work you’ve handed over, while the rest stays yours.

Lever one: the check

Every time you send an AI agent a request, it’s already running a loop of its own before you see anything: gathering context, taking an action, checking its own work, trying again if it needs to, then handing you a result (Anthropic, “Getting started with loops,” 2026). Anthropic’s own team calls this the simplest loop there is. You’re still the real check, reading the result before it goes anywhere, and the first thing worth handing off isn’t more freedom.

It’s a better check. If you’ve ever written down exactly what “good” looks like once, instead of explaining it fresh every time, you’ve already started down this path: an SOP a machine can read is what lets an agent verify its own draft against something concrete before it reaches you, rather than just reporting that it’s finished. A proposal-drafting agent that checks its own numbers against your actual pricing sheet before showing you a final version is running a real check. One that just says “done” is not, whatever it claims.

That distinction matters more than it sounds like it should, and Anthropic has the numbers to show why. Across real use of Claude Code, people approve 97% of the permission prompts the tool shows them, a rate high enough that a lot of that approving is reflexive rather than reviewed (Anthropic, “Auto mode is now the default in Claude Code,” 2026). In a controlled study with 1,053 paid testers, a single dangerous command was quietly swapped into one approval prompt partway through each session. The testers caught it 13.6% of the time. Anthropic’s own automated check, built specifically to catch that kind of action, caught it 89% of the time. That result is specific to Claude Code’s own coding prompts, not a universal verdict on human judgment, but the lesson travels past code: a person present is not the same thing as a real check, and a check built to catch one particular problem will catch it far more reliably than a habit of clicking approve.

Lever two: the stop condition

Sometimes one pass isn’t enough. You want an agent to keep working until a result actually clears a bar you can name, not until it decides on its own that it’s tried enough. That’s the second lever: instead of judging “good enough” yourself each time, you write the bar down in advance, and a separate evaluator model checks what came back against it, sending the work back for another pass when it falls short (Anthropic, “Getting started with loops,” 2026).

The bar has to be something the agent can actually demonstrate, not a feeling. “Make this deck better” gives the evaluator nothing to check against. “Revise this deck until every slide states one number with its source, and stop after five passes” gives it something real to hold the draft up to, whether the subject is a sales deck, a set of interview questions, or a first pass at a research summary. The stop-after-five clause matters as much as the bar itself. A loop with no ceiling on attempts can burn real time and real cost chasing a target it was never quite going to reach.

Lever three: the trigger

The first two levers still start when you start them. The third hands off the starting gun too. Instead of you deciding today is the day to check something, the agent checks on its own, on a schedule or the moment something specific changes outside the conversation (Anthropic, “Getting started with loops,” 2026). A recurring competitor price check every morning. A shared inbox scanned every half hour for one particular kind of request. A project tracker watched for a status nobody’s supposed to leave sitting.

This lever runs on your own machine by default, which means it stops the moment your session does. Moving it to run continuously in the background, even while your laptop is closed, is a further step, not an automatic extension of the same permission.

Lever four: the target

The first three levers all still point at something specific: this proposal, this deck, this inbox. The fourth is the real jump, and it’s the one worth slowing down for. Instead of naming a target each time, you hand over a standing area of responsibility, no specific item attached, and the agent decides what counts as work the moment it shows up (Anthropic, “Getting started with loops,” 2026). Anthropic’s own name for this fourth handoff is “the prompt,” and that word choice is the tell: it’s the assignment itself changing hands, past the schedule. A feedback channel that gets triaged continuously is the shape of it, rather than one ticket you pointed at.

That isn’t further along the same slider as the first three levers. It’s a different kind of authority, and it’s the one lever worth pairing with a separate, explicit answer to a different question: who still gets to step in and overrule the agent, and on what grounds. A three-axis way to answer that is worth reading before you pull this lever, not after.

The register worth keeping

Before handing an AI agent the next piece of recurring work, name which one of these four you’re actually ready to give up, and write down, honestly, what the other three still require of you.

Lever	What you’d be handing off	Still yours until you decide otherwise
The check	A written standard the agent verifies against	Reading the result yourself
The stop condition	A bar it can prove it cleared	Deciding “good enough” each time
The trigger	A schedule or an event that starts the work	Starting it yourself
The target	A standing area, not a named item	Pointing at the specific thing each time

The fourth lever is the tempting one to reach for first, since a standing mandate is the version of AI autonomy people tend to picture when they imagine handing over real control. The first lever is the one that actually decides whether any of the rest is safe to hand off at all. Get the check right, and the other three are a matter of degree. Get it wrong, and a wider mandate just means the same mistake, made faster, with nobody reading closely enough to catch it.