Hermes deleted 365,463 lines of its own code against a test suite that never changed

The first agent left the source files alone. It measured the codebase, split it into 36 non-overlapping groups, wrote the work tickets from the standing goal, read the workers' reports, merged branches and ran the checks ^[25][29]. Each worker got a git worktree, a separate checkout where two agents cannot edit the same file ^[26]. Some workers delegated again, so the tree ran three levels deep below the first agent ^[28]. Every ticket named three things: the code to simplify, the interfaces to preserve, and the checks that had to pass before a commit ^[27].

The objective stayed live through a `/goal` command, which holds the goal and prompts Hermes to continue at the point where it would otherwise stop ^[19]. In Nokka's Thai rendering, the prompt asks for at least 30 percent off total lines, god files broken up, shared helpers merged, less if-if-if-else routing, and then tells the agent to finish everything with no excuses and without waiting on his decisions ^[20]. Nokka wrote that this instruction is the key, because jobs with hundreds of subagents tend to die at the moment one of them stops to ask ^[33].

Divide the main run's bill across its subagents and each one cost about $13.86 ^[3]. The deletion came to 365,463 lines of non-test Python ^[1], or roughly 5.3 cents a line ^[2].

The test suite stayed as it was. In Nokka's translation, the gist the team published says the suite that guards behaviour is the same suite ^[18]. Two counts went up: functions by 25 percent and Python files by 34.5 percent, because the big ones were broken apart ^[16]. If both counts cover the same file set, mean lines per Python file fall to about 49 percent of what they were ^[5].

The two ends of the team's manual estimate differ by a factor of twelve ^[6], so the multiple depends on which end you take. Nokka's account divides out about 93 times at the top and about 7.8 times at the bottom ^[7], and Nous Research titled the post around turning $19,000 of tokens into $1.8 million of value ^[8]. The token bill was paid; the upper bound is the team's own estimate of what it would have spent on a small team working for as long as two years ^[6].

For the ratio to transfer, a codebase has to start roughly where this one started: a single `gateway/run.py` of 34,847 lines ^[10], a longest function of 7,310 lines that came out at 566 ^[14], peak complexity of 1,075 falling to 84 ^[15], and 1,338 functions over 100 lines coming down to 139 ^[13]. It also needs the skill file. `hermes-agent-dev` grew out of daily work, with Hermes recording the reusable lesson itself whenever the team settled on a method or corrected one of its mistakes ^[21], in Markdown files that it writes and edits mid-run ^[22]. One instruction in it tells the agent to reproduce a failure on `origin/main` HEAD in a clean environment before assuming its own work caused it ^[23]. Months of ordinary sessions went into that file before the September run, and the team hands it to other engineers to install in their own Hermes ^[24].

Nokka wrote that the four places where the team set out its own limits matter more than the 93 times figure ^[30]. The account, published on 16 September 2026 and itself drafted by a model running through Hermes Agent before Nokka reviewed it ^[31], breaks off inside the description of interface checking, before it reaches those four ^[32]. On the record as it stands, the checks on a 365,463-line deletion were the unchanged suite and a per-ticket interface comparison ^[17][27].