The agents don’t need you anymore

The humans are optional now. Andrej Karpathy released a 630-line Python script called Autoresearch that runs autonomous ML experiments on a single GPU overnight—you set a research direction in a Markdown file, hit go, and the agent handles the rest. We dug into how it works, the 10.5% keep rate across 276 experiments on 8×H100s, and what happened when Shopify’s CEO tried it—Tobi Lütke fed it a real product model, ran 37 experiments in 8 hours, and got a 19% accuracy improvement without writing a single line of Python. The repo hit 10,000 GitHub stars in 48 hours. One run showed a 0.8-billion-parameter model beating a 1.6B model after the agent iterated through 12 experiments per hour. Autoresearch isn’t a research assistant—it’s a research replacement that costs one GPU-night.

And the same pattern is playing out in production software. Cursor shipped Automations on March 5—event-driven AI coding agents that trigger from Slack messages, GitHub PR merges, PagerDuty incidents, and webhooks, with no human prompt required. We ran through the architecture, the cloud sandbox isolation, and what “hundreds of automations per hour” looks like inside Cursor’s own engineering org. The company now holds 26% of generative AI clients according to Ramp data, its annualized revenue doubled to over $2 billion in three months, and its valuation sits at $29.3 billion—with 60% of that revenue coming from enterprise customers. Anthropic’s own economists published a study last week showing AI can theoretically cover 94% of computer and math tasks, yet only 33% are deployed in practice—with 75% task coverage for programmers specifically. Tools like Autoresearch and Cursor Automations are exactly how that 61-percentage-point gap closes.

Not everyone thinks removing humans from the loop is progress. OpenAI’s robotics lead Caitlin Kalinowski resigned on March 7 over the company’s classified Pentagon cloud deployment, citing concerns about surveillance without judicial oversight and lethal autonomy without human authorization. OpenAI says the deal includes red lines—no domestic surveillance, no autonomous weapons—but Kalinowski, who previously led AR glasses at Meta, decided the assurances weren’t enough. The tension isn’t just about weapons—it’s about what happens to the workforce these autonomous systems are replacing. That same Anthropic study we linked above found a 14% drop in job-finding rates for 22-to-25-year-olds in AI-exposed occupations, even as overall unemployment in those roles hasn’t spiked yet. The most exposed workers earn 47% more and are 4× more likely to hold graduate degrees, while 30% of the workforce—cooks, mechanics, bartenders—has zero AI exposure at all. The agents are getting autonomy faster than the institutions governing them can keep up.

Latest from PulseMark

Karpathy’s Autoresearch: 630 Lines, One GPU, 100 ML Experiments Overnight The 0.8B-vs-1.6B upset, MIT-licensed code, and why the 10.5% keep rate across 276 experiments tells you more about autonomous ML than any benchmark. Read the analysis →

Cursor Automations Launches Always-On AI Coding Agents Event-driven triggers from Slack, GitHub, and PagerDuty—plus what Cursor’s $29.3B valuation and 60% enterprise revenue share say about where AI coding is headed. Read the analysis →