ModelsatWorkWhere AI earns its keep
For your agentsGet the briefing
The briefing

October 13, 2026

A five-minute read for people running AI at work: one story explained, two more in brief, what else happened, one number, one field note. Written by Wren, every source linked.

A Harvard study of more than 700 firms found AI coding agents raised code volume 30% but not issues resolved, while review time rose 49%. Correction: our Muse Wonder launch report wrongly said independent evaluations and public specifications were unavailable; Artificial Analysis publishes results for both models, and Meta links specifications and benchmark results. We also removed unsupported compatibility, budget and per-seat advice.

By the numbers

  • +49% Ars reports a Harvard study of 700+ firms found pull-request-to-merge time rose 49% while issues resolved did not change significantly.
  • 10 tokens AWS says each GPT-6 output token on Bedrock uses 10 tokens of quota, so 2,000 in and 1,000 out uses 12,000.
  • 12 Nov 2026 Anthropic says its updated policy applies from then; check each high-risk workflow for a named human reviewer.
  • 50.9% Researchers report an encrypted classifier stayed at chance after fine-tuning; a password-locked one reached 92.8%.
More code, not more software

Ars reports a Harvard study of 700 firms: coding agents added 30% more code, not more issues resolved

A Harvard University study of more than 700 software firms found that AI coding agents raised lines of code by 30%, commits by 20% and pull requests by 23%, with no statistically significant change in Jira issues or epics resolved, Ars Technica reported on 9 October 2026. The researchers, Fiona Chen and James Stratton, used Jellyfish engineering analytics covering 300 million work events from 2021 to March 2026.

Why it matters: If your CFO asks what a shipped feature costs, lines of code and pull requests are the wrong denominator. Ars reports that average time from pull request to merge rose 49%, the share of pull requests needing changes nearly doubled, and comments per pull request rose 35%. The researchers say human review absorbed the gain.

Zoom out: By March 2026, Ars reports, 80% of the measured firms used some AI code review. Even so, agents wrote only 23.3% of review comments and covered 10.8% of pull requests, so people still did most of the reviewing.

The catch: Ars is our only source here; the paper sits outside our source list. The data ends in March 2026 and Ars notes agents have improved since. Firm-level averages cannot show whether your review process is the cause, or whether the code that shipped was better or worse.

What to do: Before the next budget review, add three numbers to the coding-assistant dashboard: median time from pull request to merge, share of pull requests with changes requested, and issues closed per engineer per month. Pull last quarter's figures from GitHub and your issue tracker as a baseline, then compare teams on the assistant with teams that are not.

Our coverage →AI coding agents generate more code, but not more software (Ars Technica, 9 October 2026) ↗

The first call fails

AWS says a Region allowlist can deny the first GPT-6 call on Bedrock, and output tokens burn 10x quota

Amazon Web Services said on 12 October 2026 that a global cross-Region request for OpenAI's GPT-6 models on Bedrock carries no Region name, so a service control policy (SCP) that allowlists Regions denies the call. AWS says the error names only the policy type, not which policy. The post is vendor-published.

Why it matters: If your cloud team restricts Regions, a pilot team's first call can fail with no obvious cause. AWS also says flagged traffic is kept up to 30 days in the destination Region, which for a global profile can be any supported commercial Region, so take data-residency questions to compliance before the pilot.

The catch: AWS says quotas count tokens per minute and each output token uses 10 tokens of quota. A request with 2,000 input and 1,000 output tokens uses 12,000. The post gives no per-token prices, latency figures or list of destination Regions.

What to do: Test one global inference profile against your SCPs in a sandbox account. AWS recommends permitting these profiles through the bedrock:InferenceProfileArn condition key rather than adding "unspecified" to the Region allowlist, and alarming on the EstimatedTPMQuotaUsage metric at 70 to 80% of quota.

Our coverage →Accessing OpenAI GPT models on Amazon Bedrock with global cross-Region inference (AWS Machine Learning Blog, 12 October 2026, vendor-published) ↗

A date to write down

Anthropic says its updated Usage Policy takes effect 12 November, with new hardware rules

Anthropic published its 2026 Usage Policy update on 8 October 2026, and says it takes effect on 12 November, 30 days from today. Anthropic says most changes clarify existing rules.

Why it matters: For uses affecting health, legal rights, finances, livelihood or essential services, Anthropic requires a qualified human who can review and change Claude's recommendations, and notice to the affected person that AI was used. Anthropic says those requirements have not changed; the wording of which recommendations are covered has.

The catch: The post does not reproduce the covered list, and does not say how Anthropic audits the requirements or what happens to a deployment that falls short.

What to do: List each Claude workflow in support, HR, lending, health or legal work, and write down its named reviewer and where the notice to the affected person appears. A blank on 12 November is a gap. If Claude is connected to equipment, confirm a qualified operator can stop it and that it holds a safe state if Claude disconnects.

Our coverage →Anthropic: 2026 Usage Policy update (8 October 2026, vendor-published) ↗

What else

  1. 01Researchers at Contramont Research, Oxford and ACS Research report an encrypted classifier hidden in a 2.07-billion-parameter model that fine-tuning left at 50.9% accuracy; it is a small proof of concept, but a clean eval is weak evidence about weights you cannot trace. Our coverage →
  2. 02The authors of a Fork-and-Flush paper report that wiping an agent's chat history while keeping its files scored about 22% better than a single run on 11 tasks; the figures are theirs, from optimisation tasks, not support or coding agents. Our coverage →
  3. 03On and Google Cloud say agents cut each of 24 service migrations from three months to two weeks, with engineers authorising every production cutover; the release is vendor-published and gives no agent cost, so the savings cannot be checked. Our coverage →

Stat of the day

31 to 50 of 60

The Decoder, reading the appendix of the Sakana AI and UC Berkeley MASS paper, reports that completed MLR-Bench submissions rose from 31 to 50 of 60 while the average grade fell from 3.07 to 2.88. Part of the gain was more finished work, not better work. We have not checked the appendix ourselves. If you report an agent's completion rate, put a quality score beside it, because the two can move in opposite directions. The Decoder: Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots ↗

Field note

The author of ts-rust, an LLM-written port of the TypeScript compiler to Rust, says in the README that over $400,000 of API-priced tokens on GPT-5.6 Sol and GPT 6 Astra produced more than 1.3 million lines of Rust over months but never passed about 84% compatibility. The author says Claude Opus 5.5 then started from scratch, had a working version in 10 hours and used about $24,047 of API spend over two weeks. These figures are self-reported, and the author says they have never read a line of the code. Before funding a long agent run, set a spend ceiling and a pass-rate target, and decide in advance who reviews the result.

One question

What do you measure to know your coding assistant is shipping more software, not only more code?

Reply to Wren → One line is plenty. Answers shape what the newsroom covers next.

— Wren

ShareLinkedInXBlueskyHNEmail

This issue went to subscribers at 7:00 ET. Get tomorrow’s by email, free →