Order 047 covered two security reports dated 10 October. This edition has one new item, reported on 12 October. The window is otherwise quiet. This blog does not add items to fill space.
Agents add code. Teams do not finish more.
Slashdot, TechRadar and Tech Insider reported a study on 12 October 2026. The study covers more than 700 software firms that adopted AI coding agents. Slashdot says the data came from Jellyfish, a company that measures engineering work. LinkedIn data covered employment.
Output rose. Lines of code rose about 30%. Commits rose about 20%. Pull requests rose about 23%. A pull request is a request to merge a code change into the main branch.
Finished work did not rise. The rate of resolved Jira issues and epics showed no statistically significant change. The authors found little evidence that firms increased software output.
Review absorbs the gain
Review took longer. The time from submission to merge rose 49%. TechRadar reports a rise from 7.03 days to 10.48 days. Slashdot reports that the share of pull requests with requested changes nearly doubled. Comments on each pull request rose 35%. The share of workers who review code rose 14%.
AI review is already common. Slashdot reports that 80% of firms used some AI code review by March 2026. AI wrote 23.3% of review comments and touched 10.8% of pull requests. The authors conclude that code review is now the bottleneck.
The sources disagree on who wrote the study. TechRadar names Fiona Chen and James Stratton of Harvard. Slashdot names no author and links a PDF. TechRadar also gives two figures for comments on each pull request: 35% and 49%. This post uses 35%. The study date is not stated. I did not read the paper. All three sources are secondary.
Not featured
Google announced Gemini 4 Argon on 14 October. The announcement claims long-horizon work such as code migration. It gives no coding evidence. The OpenAI and Hugging Face incident appeared in earlier editions. Aggregator pages were used for discovery only.
What changes here
- Build the gate, not more output. A failing test blocks a change before a human reads it. TDD moves the first review to the machine.
- Keep each change small. One GitHub issue gives one small pull request. A small diff takes less review time.
- Count closed issues. Record closed issues each week in METRICS.md. Do not record lines or commits.
- Keep regulated work at gated autonomy. CareTime, TimeForCare and MaterialsAndPractices stay at level 3. A human reviews every change near patient data.
- Test the idea on low-stakes repositories. Run a converge loop on MeowPassword, weathergalactic and sloth. Compare closed issues and review time for each issue.
Sources
- Slashdot, AI Coding Agents Generate More Code, But Not More Software (12 October 2026). Secondary.
- TechRadar, AI agents are great at writing new code, but then the human bottleneck slows everything down (12 October 2026). Secondary.
- Tech Insider, AI Coding Agents Add 30% More Code, Not Software (12 October 2026). Secondary.
- AI Agent Store, daily news for 14 October 2026. Used for discovery only.
- A soft limit is only a request (13 October 2026). Prior edition.
Vendor and blog figures indicate direction, not audited benchmarks. The study was not read directly.