Agents on the Remote Labor Index Jump from 2.5% to 41.3% in a Year

Most benchmarks for agents look alike: solve a task from a repository, answer questions, run through a store simulation. All of that is far from what people actually pay money for. The Remote Labor Index is built differently, which is why the figure from a recent Telegram post deserves attention.

What the index actually measures

The Remote Labor Index (RLI) was created by Scale AI and the Center for AI Safety in autumn 2025. The idea is simple: take real projects that clients posted on freelance marketplaces, hand them to an agent, and show the result to the person who paid. If the client says "this is on par with or better than a human," the task counts. The final metric is the share of such projects.

In the first version, the best agent scored 2.5%. Now, according to the post, the figure has grown to 41.3%. That is roughly a 16-fold increase in a year.

The mix of tasks matters even more. The index covers 23 categories: 3D and CAD, architecture, graphic design, video and animation, audio and much more. Code is just one of them. For a product manager or founder, this is the key signal: agents are moving beyond programming, where they are usually evaluated.

How to read the 41.3% figure

Imagine a freelance order: "make a 30-second animated clip for a landing page." Before, the agent produced something the client would immediately throw in the trash. Now, in four out of ten such cases, the result is accepted as usable. That doesn't mean the agent has replaced a studio. It means that on every second or third order, there is no longer any reason to go straight to a human.

But the figure comes with caveats, and the post doesn't address them:

• It is unclear which model or system scored 41.3%. It could be a single agent, or the best result among many.
• The post has no link to the primary source, only to the author's Telegram channel. The results table should be checked on the index's own website.
• The client's rating remains subjective. "As good as a human" on a one-off project does not guarantee consistency across a stream of a hundred orders.
• The remaining roughly 59% of tasks are still not closed by the agent. Often these are the most complex and expensive projects.

The bottom line: the trend is real, but the exact scale is better verified against the source than a retelling.

Unemployment or not

The post's author jokes about unemployment, and the joke stings. But in practice, such metrics rarely turn into "a human was replaced entirely." More often, the structure of work changes: the agent produces a first draft, and a human reviews and refines it. For freelancing, this means pressure on prices for simple orders and growing demand for people who can accept results and set tasks.

What to do in practice

Developers and product managers should do three things.

1. Take your typical non-code tasks: mockups, illustrations, short videos, voiceovers, simple CAD models. Run them through modern agents and compare the output with human work against your own quality criteria.
2. Build an acceptance process. If the agent meets the standard in 40% of cases, value is created by whoever quickly filters out the other 60% and doesn't waste time on them.
3. Track the index as a barometer. For hiring and budget planning, the trend by category is more useful than a single overall number.

The main takeaway: a year ago you could safely say that agents can't handle real work. Now that claim needs to be checked against specific tasks.