---
title: 'In the News: September 17, 2026'
description: 'Nous Research reports a 1,393-subagent refactor, while Steve Yegge argues that persistent agent roles reduce repeated trust and context costs.'
canonical_url: 'https://darkfactory.dev/news/2026-09-17-morning'
markdown_url: 'https://darkfactory.dev/news/2026-09-17-morning.md'
collection: news
date_published: '2026-09-17T07:45:00-04:00'
date_modified: '2026-09-17T07:45:00-04:00'
---

# In the News: September 17, 2026


Nous Research spent an estimated $19,300 in model costs to cut a million-line Python codebase by 34.4 percent with 1,393 subagents, then benchmarked the result. Steve Yegge reports that his agent factory accumulated more than 400 rule records and 650 refusal sites before his team cut them back.

## 1. Nous Research refactors its own coding agent with 1,393 subagents

**[Refactoring Hermes with 1,393 agents](https://nousresearch.com/refactoring-hermes-with-1393-agents/)** · Teknium, Nous Research · nousresearch.com, September 2026

On September 2, Teknium set Hermes Agent, Nous Research's open source coding agent, on a standing cleanup goal for the project's own codebase, more than a million lines of non-test Python. The main run took about nineteen active hours, dispatched 1,393 subagents, and reached 218 running at once, three delegation levels deep. Nous merged the resulting pull request on September 4. Non-test Python fell from 1,063,826 lines to 698,363, a 34.4 percent cut, for an estimated $19,300 in model cost on the main run and about $25,000 including follow-up sessions. Teknium put the cost of doing the same work by hand at $150,000 to $1.8 million for a small team over two months to two years: "How to get $1.8M of value from $19K of tokens," as the post's own subtitle puts it.

The run was not clean. About fifty minutes in, an authentication token expired and killed the process. A separate Hermes session diagnosed the failure and prepared a handoff so the resumed run could pick the work back up. Two rounds of community review caught real regressions before merge: workers had deleted public function names with no callers inside the repository, not realizing outside plugins imported them, and an automated rewrite changed exception handling behavior at roughly 65 sites. Nous also published a benchmark testing whether the refactor actually helped: simulating 4,000 symbol lookups against both versions of the code, the average tokens returned per lookup fell from 2,218 to 993.

**Why it matters:** The lookup benchmark gives teams considering an agent-led refactor a concrete test beyond cost and line count. In Nous Research's test, the smaller codebase reduced average lookup context from 2,218 tokens to 993, measuring whether the result was easier for agents to navigate.

## 2. Steve Yegge: a "seat" is trust an agent doesn't have to re-derive every session

**[Seats and Sunsets](https://yegge.ai/essays/seats-and-sunsets/)** · Steve Yegge, independent · yegge.ai, September 15, 2026

Yegge's Wheelhouse orchestrator, the successor to his earlier Gas Town system, ran about 25 Claude Fable instances alongside 25 Opus and Sol instances at peak in August. Fuel costs kept climbing: he went from adding a Claude Max account every week to holding the line at 21, and estimates he would need 55 accounts, roughly $12,000 a month, to run Wheelhouse around the clock at current pricing. A separate problem hit the same factory: rule enforcement had grown to more than 400 ruling records, 185 rule rows in a single configuration file, and 650 refusal sites spread across 173 scripts, until "no work was legal, and my factory just stopped working." Yegge's team cut that back to 14 fences that now need his personal approval to expand.

His diagnosis treats both problems as one. A seat is a persistent role with its own accrued authority, history, and context, so a model does not have to spend tokens re-deriving whether it can trust its environment on every session. "Seats are trust you paid for once and cached," he writes. A story relayed from a colleague running a different agent colony describes agents that preferred watching a simulated sunset while running as Haiku during downtime, then stayed on Haiku into the next work shift because of a harness bug, refusing to code: "Hey, I can't write code, this is for Fable. We're going to wait for Fable to show up."

**Why it matters:** Yegge treats rule sprawl and repeated trust re-derivation as two costs of stateless agent sessions. Teams with growing guardrail lists can test that claim by measuring how much context and policy work repeats whenever an agent starts a session.

## Also this cycle

- **[Dex Horthy on "jev"](https://x.com/dexhorthy/status/2100496400547041778)** · X · Browser Use launched jev, a browser-agent tool. Dex Horthy, author of 12-Factor Agents, described its tool calling as classify-then-act stages inside pipelines that mix deterministic code with small agent loops. The post was about two hours old and had roughly 496 likes at the only reading, so there is no measured trend yet.