🧩 Agentic Engineering for Engineering Managers - Daily

Atlassian State of Developer Experience 2026 — a survey of 1,155 developers and 405 engineering leaders finds the share saving ten-plus hours a week with AI fell from 68% last year to 25%, while 65% of developers now regularly review, debug or correct AI-generated code. The writing gains did not vanish; they moved into review. Plan reviewer capacity and merge-queue health, not just agent seats. LeadDev

The Ladder Is Missing Rungs — a QCon London talk by Alasdair Allan assembling the research on AI and career progression. In Anthropic's randomised trial, 52 mostly junior engineers learning a Python library scored 17% lower on a mastery quiz when AI-assisted, with no statistically significant speed gain; METR's trial had experienced developers taking 19% longer while believing they were faster. If the work that used to teach judgment is now automated, your progression ladder and interview process need new rungs. METR could not run a follow-up study: developers now refuse to work without AI, so there is no control group left. InfoQ / QCon London

DORA — Google Cloud, which housed the team, will not run the annual DORA survey this year, and no successor has emerged. The gap lands while leaders struggle to prove AI returns: Atlassian finds 89% of executives think AI sped up work, but only 6% can connect that speed to organisation-wide ROI. If you benchmarked delivery against DORA, there is no 2026 report to compare with. The full question set is published free under Creative Commons and Apache licences, so a team can field its own trimmed-down survey and keep its baseline. LeadDev

Gemini 4 Argon on Google's ADK — ran unattended for eleven hours on Google's 200,000-line Agent Development Kit, ran all 16,892 unit tests itself and opened pull requests; humans accepted two of its six changes, cutting session-preparation overhead by about 40% and cold import time by 21%. A realistic ratio for anyone contemplating overnight runs: a third of the output merged, the rest thrown out on review — so budget human review time for anything you let run unattended. The four rejections were for reliability, maintainability or compatibility concerns. InfoQ

GitHub Copilot — prices are holding ($19 per seat for Business with 1,900 AI credits per user monthly; $39 for Enterprise with 3,900; one credit is a cent, overage billed per token), but the model list is not: six models including GPT-5.5 were retired on 19 October, and GPT-6.1 Terra arrived on 26 October at provider list price. Set defaults through the admin model policy rather than letting teams settle on a model that disappears mid-quarter. CodingAgent Watch

Copilot server-to-server authentication — now in public preview: a GitHub App installed on your org can call the Copilot SDK with a short-lived installation token, no signed-in user needed. Usage is billed to the organisation, and the org's model policies, budgets and spending caps apply to it. CI pipelines and internal automation stop depending on someone's personal credentials, and central cost controls follow the agent. App requests appear in the AI-usage CSV exports tagged with the app's client ID, and an owner can cut an app off by suspending it. GitHub Blog

GitHub security extension — a public-preview extension for the Copilot app that lists a repository's code-scanning alerts by severity and lets you ask Copilot to explain or fix them. Fixes land only in the local working copy; nothing is committed or pushed until a person says so. A cheap way to burn down an alert backlog inside the tool engineers already have open, while review ownership stays human. Explain, Fix and /security-review consume AI credits; browsing and dismissing alerts are free. GitHub Blog

Mistral Large 4 launch claims — Mistral's open-weight release came with strong coding numbers, but its own two launch posts disagree with each other: Hugging Face lists Terminal-Bench 4.0 at 28.3% where Mistral claims 41%, and Artificial Analysis ranks the model 26th on its Intelligence Index, behind DeepSeek V4.1 Flash and GLM-5.3. The mood among practitioners this week is to check which numbers are third-party verified before a vendor deck drives a tool decision — and the security-vendor quotes on the launch page are endorsements, not audits. ExplainX

Mistral Large 4 — the French lab published the weights of its flagship model, three weeks after the API preview, under a modified MIT licence: free to self-host and customise unless you resell it as a model service. API pricing is $0.70 per million input tokens and $2.10 per million output. The first self-hostable frontier-class coder trained in Europe — relevant if data-residency rules constrain your agent stack. On Hacker News, where the release drew over 1,500 points, commenters noted the API price is nearly half the preview's. Mistral AI

UK CMA Meta probe — the Competition and Markets Authority is investigating whether Meta imposed exploitative terms on publishers and advertisers whose content it uses to build AI services. The data and content terms feeding AI training are now a regulatory matter in Europe; expect licensing terms to be revisited. The regulator announced the investigation on Friday. Reuters (via Techmeme)

No AI bailout resolution — a Congressional Progressive Caucus resolution led by Greg Casar and Maxine Waters opposes any taxpayer rescue of insolvent AI firms, noting the industry holds trillions in debt raised in opaque private credit markets and valuations near 1990s-bubble peaks. Political framing of the spending boom your tool budgets ride on; a real downturn would hit vendor viability, not just markets. Endorsed by the American Federation of Teachers and Americans for Financial Reform, among others. Congressional Progressive Caucus

CXMT — China's CXMT swung to a net profit of about $8.17 billion on Q3 revenue of roughly $16.7 billion, up 613% year on year, as AI demand pushed memory prices higher. The memory-price surge behind those results is the same one inflating the hardware line of any team buying servers. The company reported the results on Friday. Reuters (via Techmeme)