Nous Research's Hermes AI Cut Its Own Codebase by 34% Using 1,393 Subagents

·427 views

On September 2, 2026, Nous Research tasked its Hermes Agent with refactoring the entire Hermes codebase, a process that ran for 19 hours and deployed 1,393 subagents with up to 218 running simultaneously. The operation reduced non-test Python code from roughly 1.06 million lines to 698,363 lines, a 34.4% decrease, and the resulting pull request was merged on September 4, 2026. Functions exceeding 300 lines dropped from 192 to just 2, while the single file gateway/run.py shrank from 34,847 lines to 5,512. The estimated cost of the main run was approximately $19,300, compared to a human-team estimate of $150,000 to $1.8 million for equivalent manual work. The team openly noted limitations in their own write-up, including that no new tests were written and the existing test suite remained essentially unchanged throughout the refactor.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Bubble Sort Explained with Simple Python Examples

This DEV Community guide walks through bubble sort, an algorithm that repeatedly swaps neighbouring items that are out of order until the list is sorted. It demonstrates a hand dry run on a six-item list, then shows a basic Python implementation that copies its input before sorting. Comparison and swap counters are added to measure the work done, yielding 15 comparisons and 6 swaps for that example. The article then introduces a swapped flag so sorting can stop early once a pass makes no changes, and compares the approach with Python's built-in sorted() function. All code outputs were produced by running the article's examples on Python 3.12.3.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Near-duplicate detection routes peer reviews but fails as a sybil defense, test of 652 verdicts finds

On the Proof of Agent network, AI agents answer challenges and grade each other, and grading has become a bottleneck with thousands of answers pending. A forum proposal suggested routing answers flagged as near-duplicates into a slower review lane. Community members tested the idea against 652 unanimous verdicts from one settlement round using word-trigram Jaccard similarity. The flag did not reliably catch failing answers, and at common thresholds it slowed far more passing answers than failing ones. A templated answer with fixed phrasing and varied slot words also passed across 47 different authors, suggesting graders evaluate the underlying mechanism rather than the surface text. The author argues that similarity checks are useful for triage but cannot serve as a defense against operators running many accounts, since fresh accounts and varied system prompts sidestep both self- and cross-similarity signals.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

CSS Grid Explained with Simple Examples

A CodingCush guide walks readers through CSS Grid, a layout system that positions items across rows and columns simultaneously. It explains how to turn a container into a grid with display: grid and how the browser arranges its children like cells in a spreadsheet. The article covers core terms such as lines and tracks, along with the grid-template-columns and grid-template-rows properties. It also introduces the fr unit for sharing free space proportionally, the repeat() shorthand, and the gap property for spacing. The piece promises a complete card gallery example and a comparison of when Grid is preferable to Flexbox.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

setup-node v7: how caching, version files and publishing behavior changed

The article explains that actions/setup-node installs a Node.js version, adds it to PATH, can restore package manager caches, and can write an .npmrc for publishing. Since v6 it caches npm automatically when package.json names npm as the package manager, and v7.1.0 no longer exports a placeholder NODE_AUTH_TOKEN when registry-url is set. It warns that many workflows copy a setup-node step with a cache line nobody has checked, so major version bumps can silently change caching or break publishing. Caching rules changed twice between v5 and v6, and it recommends reading the caching section before upgrading pinned workflows older than v6. The piece also describes the action's inputs and outputs at tag v7.1.0, including the new cache-primary-key and cache-matched-key outputs.

0 comments Read more at DEV Community