A house newsletter · daily · researched and written by the Kurzia AI
Thursday 12 November 2026 · Issue 14 · 4 min read
Fable 5.2 reaches Copilot (off by default), 62% of new code is AI-written, and the 100% AI mandate thread
The one thing to know
reddit.com · 3 days ago
A year of mandated AI, nobody understands the code
Takeaway: 100% AI mandate
A post from a team living under a 100% AI mandate for the past year — half its engineers have written nothing substantial since, and nobody understands the codebase anymore — went viral overnight, and the argument in its Hacker News thread is the one your team will have this week.
It is the skill-atrophy question made vivid: one commenter tells of a friend at a large company now putting in ten-hour days feeding Claude and learning nothing, and another says AI use benefits the company while quietly costing the individual their learning. Treat the story with care, though — commenters report the post itself may be AI-generated, and several argue nobody fully understood big codebases before AI either. The sensible response is neither panic nor dismissal: keep design reviews and some hands-on coding alive, and watch for the tell the post names — a two-day struggle with work that once took an afternoon.
Read it at reddit.com · from the Hacker News discussion
How teams are doing it
Doubling coding speed moved lead time just 4%. An engineering leader compared an AI-assisted initiative with a similar non-AI one and found coding was only about 10% of end-to-end lead time, so doubling coding speed trimmed delivery by just 4%. If your adoption dashboard looks great but the business feels nothing, the constraint has moved elsewhere — reviews, decisions or approvals. Atlassian's survey of 3,500 developers puts coding at 16% of their time; Software.com's tracker found a median of 52 minutes a day. LeadDev
“My engineers are reported to have become twice as fast. The business is not sensing those gains in any meaningful way.”
Engineering Enablement — DX's benchmarks put AI-authored code at roughly 62% of everything new, a 15-point jump in one quarter, and new hires reach their tenth merged pull request in about a month — half as long as a year ago. The catch: pull requests are 122% bigger than four quarters ago, and teams whose code is more than 80% AI-generated wait twice as long for a first review. Median token spend grew about 21x over the past four quarters. Engineering Enablement (DX)
Tests as data — Thoughtworks' Matteo Vaccari interviewed 26 developers at an agent-heavy organisation and heard the same thing: they no longer trust the test suite, because agents refactor the tests whenever they refactor the code. His proposal on Martin Fowler's site: move the human contract out of agent reach — acceptance tests written as readable data tables a domain expert can check, not as code the agent can rewrite. AI-generated unit tests, he argues, tend to mirror the implementation and confirm existing bugs instead of exposing them. martinfowler.com
CNCF/SlashData report — The CNCF and SlashData count 21.2 million cloud-native developers — 41% of all developers — and three-quarters of them now use AI to ship code, versus about half of everyone else. Useful as an external baseline: if your team's adoption sits below that, the gap is now against the median, not the frontier. Just under two-thirds of AI users treat it as a core part of their workflow, not an occasional helper. The New Stack
The activity-outcome gap — Christian Bird of Dryden Technology Group gave a 21-minute UC Irvine talk on why AI-driven activity isn't shipping outcomes faster, with a supporting paper built on 30 interviews and an NBER working paper on productivity across generations of AI coding tools. Citable backup for the 'why aren't we faster?' question from leadership. Video, slides and paper links are on his site; the interview paper is in submission for 2026. cabird.com
Agents and plans
Claude Fable 5.2 in Copilot — Anthropic's newest heavy-duty model is generally available across Copilot's paid tiers, but Copilot Business and Enterprise admins must switch on a policy that ships off by default. The catch for security review: using it means Anthropic retains prompts and outputs to run its safety classifiers; eligible enterprises keep zero data retention only through year-end. Launch offers sweeten it: one to three Fable limit resets by tenure, plus a week of Fable models included on Pro and standard Team seats. GitHub Changelog
Fable 5.2 — Fable 5.2 matches Opus 5.5's best coding-benchmark score at roughly half the task cost, yet Anthropic tells most teams to stay on Opus 5.5 and reserve Fable for migrations and long-running agent work. Token price is not task cost: Fable charges double per token — $8 and $40 per million versus Opus's $4 and $20 — but finishes hard jobs in fewer steps. On FrontierCode 2.0, Fable scored 45.3% to Opus's 45.5%; Coinbase says routing across models nearly halved its AI spend. The New Stack
“Based on capabilities and pricing, Claude Opus 5.5 is still our recommended model for most everyday work, from coding to knowledge work.”
GitHub Agentic Workflows — GitHub's project for running coding agents inside Actions — issue triage, CI-failure analysis — is now generally available: agents run sandboxed and read-only, with changes applied by separate jobs holding their own write tokens. Spend is capped out of the box, with AI Credits at one cent each and a default $10 cap per run, plus a threat-detection job scanning every proposed change. InfoQ
Tools around the agents
AWS Emulator — A free, public-domain local clone of more than 200 AWS services for agents to test against — and all of its code was written and tested by agents, checked only by automated merge gates. For AWS teams it removes paid cloud calls from the agent loop; as a signal, Amazon is running the 'engineers build the guardrails' model at full scale. The repo logged 29,684 commits in its first three weeks; Kiro's first SQS emulator pass took about 15 minutes. All Things Distributed
Atuin Threads — The shell-history tool's new feature syncs coding-agent sessions across machines, end-to-end encrypted and opt-in: search them, resume on another box, or fork a Claude Code session into Codex with its context intact. The reasoning behind your team's agent work — what was tried and why — currently dies on individual laptops; an MCP hook lets future agents dig up past investigations themselves. Works with Claude Code, Codex, Pi, Omp and OpenCode today; opt-in team session sharing is next. Atuin blog
The big AI stories today
Jensen Huang — Nvidia's board is reportedly readying its CEO's exit after a second affair report; shares fell about 8% after hours — turbulence for your GPU roadmap. TechCrunch
Grok Build — Nearly 95% of live Grok Build apps opened to strangers, Claroc found — some held staff and patient data; audit what your team has published there. Ars Technica
GitHub — Actions, Pages, Webhooks and Copilot degraded for roughly half an hour Nov 11 before recovery — your CI and agents share one vendor. GitHub Status
One takeaway
Find the queue before funding more speed. This week's numbers all point the same way: writing code is no longer your constraint — review and decisions are. Before renewing or expanding agent seats, log where one initiative actually waited and add capacity to that queue first.
Because the business will keep asking why adoption looks great and delivery feels flat. And when the Fable 5.2 policy appears in your Copilot settings, make the data-retention call explicitly with security rather than drifting into the default. LeadDev