Autonomous Agent Loops Are Real: Karpathy, Stripe, and Anthropic's Dreaming
TL;DR Three examples of AI agents running in loops without constant supervision, verified independently rather than taken on a post's word: Andrej Karpathy ran an agent through 700 experiments in two days and pulled 20 real improvements out of it. Stripe has internal agents merging 1,300 pull requests a week, under strict guardrails. Anthropic officially documents a mechanism, dreaming, where the…
Three real-world instances demonstrate AI agents running loops without constant supervision: Karpathy's autoresearch system ran 700 experiments in two days, pulling 20 genuine enhancements for training a small language model; Stripe's internal agents merged 1,300 pull requests per week under strict guardrails; Anthropic's "dreaming" mechanism involves agents re-reading their own history to extract reusable information.
While a separate claim of a 42% productivity gain from Claude Code's lead engineer lacks independent confirmation, the common thread in these verified examples is the loop architecture instead of a single prompt. Regular prompts answer once, while loops run the agent with the previous round's result until a stop condition is met—be it a set number of iterations, validated quality bar, or time limit.
This distinction stems from the architecture surrounding the loop: who initiates it, who verifies the results, and who determines when to halt. Karpathy's loop involved 700 experiments over two days, yielding 20 real optimizations for training a smaller language model; these optimizations translated to an 11% reduction in training time when applied to a larger model.
Similarly, an internal Stripe system handles about 1,300 pull requests weekly, representing a 30% increase in less than two weeks from a prior 1,000. The key difference lies in the architecture enabling this, including strict isolation, no internet or production access, and a human gate before each merge. Anthropic's "dreaming" mechanism involves agents systematically re-reading their session history to extract reusable patterns, as demonstrated in a security research demo where nine agents working in parallel spent 800 cumulative hours on a single problem, reducing the performance gap on the task by 97%.
Additionally, Anthropic mentions that Claude now writes 80 to 90% of the code for its internal tools, although this figure comes from an official company statement rather than an independent source. One unsubstantiated claim involves a knowledge-graph system attributed to Claude Code's lead engineer, with a reported 42% productivity gain on day one.
However, no independent source, technical writeup, or mention on Anthropic's official channels supports this number or the attribution, as it remains a post's speculative figure. To build a credible loop, verification must not rely on the agent's self-assessment; instead, external checks and verifiable histories ensure the loop's efficacy without accumulating errors.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.