Amjad Masad's Levels of AI Autonomy, From Language Server to Agent Fleets

The framework

Most comparisons of AI coding tools are feature lists. Masad grades them on one axis instead: how long the thing runs before a human has to step back in. He borrows the scale from cars. Lane assist and full self-driving sit at opposite ends of one ladder.

Level one is a language server, the IntelliSense that has shipped in editors for a decade. Level two is AI code completion, the Copilot generation. Level three is an agent that works ten to fifteen minutes alone. It still needs you to test the app. Level four is close to hands-off. It wants your attention now and then. Level five is the one that changes the job. At that level you spin up a thousand agents against a thousand problems and trust that most of them land.

Masad's own company is the worked example. Replit laid people off and pivoted to agents. The company bet its future on autonomous software development. Masad says that bet only became viable once models like Claude 3.5 arrived. The jump from their first agent to the second was a jump in this measure, not in features.

The ladder matters because each rung has its own failure mode. The rung you are on also decides what you build around it. Masad puts the reliability work at levels three and four on infrastructure. That means end-to-end testing and computer use. It also means a transactional file system that lets an agent fork and sample several paths before it commits to one. The scale is not a marketing chart. It is a statement about how much of your attention the tool has earned.

How to apply it

Grade the tool in front of you before you argue about which is best.

1. Time it, do not spec it. Start a real task and measure the minutes before you must intervene. That number is the level, whatever the pricing page claims.
2. Match the level to the blast radius. Level three work belongs on things you can throw away. Reach for level four only where a wrong answer costs a review, not a customer.
3. Build the guardrail the next level needs. Moving from three to four is a testing problem, not a model problem. Automated test generation and snapshots are what let an agent run unwatched.
4. Let it sample, not guess. Replit's own answer was a transactional system where agents fork and branch. Generation becomes a set of parallel attempts, not one shot.
5. Move yourself from typing to reviewing. Each level up converts your hours from production into judgment about what should exist.
6. Watch computer use. Masad names it as the missing piece holding back real automation. A tool that cannot drive a browser has a ceiling, whatever its demo shows.

The diagnostic use is sharper than the shopping use. If your team argues about which assistant to buy, the ladder turns that into a measurable question. It often shows that two tools people describe in opposite terms sit on the same rung.

When it works, when it fails

Works best when you are choosing between tools that sell themselves on features. It works when you are deciding how much review a generated change deserves. It works when you are planning what to build around an agent, not which agent to rent. It also works as a forecast. Masad's claim that generic application software trends toward zero follows from the top of this ladder.

Fails when you treat the levels as fixed properties of a product. The rung moves with the model underneath. Masad is blunt that his own predictions here have been wrong in one direction. Too slow. It also fails where unattended runtime is the wrong measure. Some changes cost little to write and everything to get wrong.

That limit is the interesting part, because it is where Masad thinks the work goes. As the levels climb, he argues the bottleneck moves from production to ideation. When anyone can build, knowing what to build becomes the scarce skill. The same logic pushes him toward generalists over specialists. Critical thinking is what still pays once execution is cheap.

The honest limit: the ladder tells you how much supervision a tool needs. It says nothing about whether the thing it builds is worth building.