AI agents are automating freelance work six times faster than most observers expected

The number that matters most in the current AI debate is not a model score on a coding benchmark or a researcher’s subjective sense of capability. It is a figure that tracks what AI agents can actually do for money, in competition with human workers, on real tasks. That figure has moved in ways that should unsettle anyone who assumed the transition to economically capable AI would be gradual enough to manage at a comfortable pace.

Nathaniel Whittemore put it directly: “The frontier has more than quadrupled in under eight months, a concrete signal of how quickly economically capable AI agents are advancing.” The benchmark he is describing is the Remote Labor Index, a measure of the fraction of freelance jobs that AI agents can complete autonomously, without human assistance.

The external record confirms his framing, and then some. The Remote Labor Index, published by Scale AI and the Center for AI Safety, registered 2.5% automation of freelance jobs at its October 2025 launch. By July 2026 that figure had risen to 16.10%, according to Scale AI and CAIS’s own published data, corroborated by coverage in arXiv, the-decoder.com, and runtimewire.com. That is a 6.4x increase across roughly eight months. Whittemore described the move as “more than quadrupled,” which is technically conservative relative to what the index actually shows.

What makes the Remote Labor Index a more useful signal than most AI benchmarks is its construction. Standard capability benchmarks measure performance on curated problems that may or may not reflect work that anyone pays for. The RLI measures something more economically direct: the share of tasks on freelance platforms that AI agents can complete at a level sufficient to satisfy a client. When that share moves from 2.5% to 16%, it is not a laboratory abstraction. It describes a shift in who, or what, is competing for a real category of paid work.

The frontier has more than quadrupled in under eight months, a concrete signal of how quickly economically capable AI agents are advancing. Nathaniel Whittemore

The speed of the movement is the part worth sitting with. Benchmark improvements in AI have often been dismissed as measuring narrow, artificial tasks, but the slope of the RLI since launch does not allow for the usual deferral. A sixfold increase in eight months, if sustained even at a fraction of that rate, would push automated freelance capacity into territory that would have seemed speculative as recently as late 2024. The index captures agents competing across a range of task types, which means the gains are not confined to a single domain where automation was already anticipated.

There is a limit to what a single benchmark can prove. The RLI measures tasks that appear on freelance platforms, which skews toward certain categories of digital work: writing, coding, data tasks, design. It does not capture the full breadth of knowledge work, and it does not say anything about physical labor or work that requires sustained client relationships. A number that looks alarming in one slice of the market may look less so once the broader population of jobs is in view.

That caveat is real, and it should be held. But the appropriate response to a caveat is precision, not dismissal. The RLI is not measuring all work. It is measuring a specific, economically grounded proxy for AI agency in the digital labor market, and that proxy has moved faster than essentially any forecast from the period when the benchmark was designed. Whittemore’s framing, that this is “a concrete signal,” is exactly right. It is not a ceiling and it is not a prediction. It is a rate of change, and the rate is high enough that the question of how quickly economically capable AI is advancing has an empirical answer for the first time.

The institutions, platforms, and policy frameworks built around the assumption that this transition would move slowly now face a benchmark that suggests otherwise. Whether the slope continues, flattens, or accelerates, the baseline has already shifted.