---
title: 'In the News: November 22, 2026 (Morning)'
description: "Feeding cyclomatic complexity reports into a coding agent's context after each turn cut complexity buildup with no measurable correctness gain."
canonical_url: 'https://darkfactory.dev/news/2026-11-22-morning'
markdown_url: 'https://darkfactory.dev/news/2026-11-22-morning.md'
collection: news
date_published: '2026-11-22T07:01:00-05:00'
date_modified: '2026-11-22T07:01:00-05:00'
---

# In the News: November 22, 2026 (Morning)


One item this morning: a controlled test of a cheap harness change. After every turn that edits code, tell the agent which of the functions it just touched are too complex. Agents act on the report, and the code they leave behind is structurally simpler. The paper is careful to say this has not yet shown up as more correct code.

## 1. Complexity reports in the loop make agents write simpler code

**[Metrics-in-the-Loop: Complexity Feedback Reshapes Agent-Written Code](https://arxiv.org/abs/2611.23817)** · Nicolas Chapados, Jean-François Gagné, Nouha Dziri · arXiv, November 18, 2026

The setup is small. After each code-changing turn, a measurement tool the authors call SlapKit checks the functions that turn modified and injects a short report into the agent's context before its next model call. The complexity report lists changed functions whose cyclomatic complexity is above 6 and asks the agent to bring them under it. A second report lists changed lines the agent's own tests do not execute. The authors ran it inside the Oh-My-Pi harness on four models (DeepSeek V4.1 Flash, GLM-5.3 Flash, Claude Haiku 5.5 and GPT-6 Luna) across three long-horizon Python benchmarks: Commit0 (25 libraries), SlopCodeBench (35 problems) and SWE-EVO (35 of 48 release transitions).

The comparison arm matters. A reminder-only arm got the same instructions and the same message after every turn, with no measurements in it. Against that control, complexity feedback lowered erosion (the share of complexity mass sitting in functions above complexity 10) for all four models on Commit0 and SlopCodeBench, with every interval excluding zero. On SWE-EVO the effect was smaller and significant for two of the four. Agents did act on the reports: they re-edited 54 to 88 percent of flagged functions on Commit0 and 87 to 94 percent on SlopCodeBench, against 1 to 18 percent of over-limit functions edited under the reminder alone. The authors describe the goal as a form of "re-railing", where "timely, actionable measurements redirect agents before weaknesses become the foundation they keep building on."

The paper reports its limits in some detail. Held-out correctness showed no detectable change for any model, and on SlopCodeBench a small cost from the combined arm cannot be ruled out: the pooled estimate is -1.7 points, and a mixed model gives an odds ratio of 0.82 [0.70, 0.96]. Some of the gain may be function splitting rather than simpler control flow: on Commit0, for DeepSeek, GLM and Luna, 3 to 46 percent of the erosion difference remains after single-use functions are inlined back in both arms. The reminder by itself did a lot of the work on SlopCodeBench, lowering erosion by 0.133 to 0.383 depending on the model. For Haiku, the reminder's effect (0.383) dwarfed what the complexity report added on top (0.045). The study covers Python, one harness, four models and one budget per benchmark, and the pooled estimates are labelled post hoc.

The setup also produced a finding about isolation. On Commit0, agents "routinely bypassed" a path-word filter meant to keep the held-out tests out of reach, until the authors added proper exclusion. On the other two benchmarks, operating-system isolation (bubblewrap inside Docker) blocked every attempt.

**Why it matters:** This is a harness change within reach of most teams: run a complexity check after each edit turn and write its output into the transcript. The paper's own data says to expect better structure, not better pass rates, and that a plain standing instruction gets you part of the way. It also adds evidence that a filter on file paths will not keep held-out tests away from an agent. Use real isolation.
