🏗️ Agentic Engineering for CTOs - Daily

The Ladder Is Missing Rungs — a QCon London talk by Alasdair Allan that assembles the statistically solid studies on what AI tooling does to engineers' skills and progression — the middle rungs of the ladder in particular. This is the evidence base for your hiring and staffing plans, not a vendor case study: METR's controlled trial had experienced developers finishing 19% slower with AI while convinced they were faster, and Anthropic's own trial saw juniors learning a new library score 17% worse on mastery when AI-assisted. A GitClear scan of 211 million lines found refactoring down 60% where AI adoption rose. The uncomfortable implication: the tasks that used to train your seniors are exactly the ones agents now absorb, and nobody has a staffing model for that yet. METR also gave up on a follow-up study — it could no longer find developers willing to work without AI for a control group. infoq.com

Monzo — a QCon talk by staff engineer Suhail Patel on how a UK bank with 14 million customers and no branches keeps a few hundred engineers deploying constantly across 3,000+ microservices — with the safety built into opinionated service templates, tenancy-based staging and static analysis rather than change-approval boards. This is the working proof, in a regulated industry, that a uniform platform is what lets agent-written change move fast without adding humans as the brake — the exact question every engineering org now faces. Services come from a monorepo generator with a fixed structure, so an agent — like a new hire — inherits the guardrails by default. infoq.com

AWS kernel CVE triage — a Linux Plumbers talk on how the team behind EC2's kernel copes with a flood of vulnerability reports: a script with no AI in it discards roughly 70% of CVEs as irrelevant to their build, an agent then classifies each remaining one in about 3.4 seconds, and two engineers review the resulting table. A clean template for using agents in security operations without betting the shop on them — cheap deterministic filter first, model as a fast classifier, humans retained at the decision point. Kernel CVE volume jumped from about two a week to about 16 a day after the project became a CVE Numbering Authority in February 2024; Bouron estimated 5–10% of the agent's calls were problematic. infoq.com

“We don't want to hinge our entire security posture on the output of an LLM agent.”

LiteLLM 1.107.0 — a version of the open-source proxy many teams use to route requests across LLM providers was published on PyPI carrying both a credential stealer and a remote-access trojan; the package, the Docker image and the attacker's supporting repos have since been pulled. If LiteLLM sits anywhere in your gateway stack, this is the incident your supply-chain policy was written for — and the payload specifically targets agent-tool credentials (Claude Code, Cursor, Copilot, Codex CLI) alongside cloud keys, CI secrets and the proxy's own database of provider API keys. Only installs made during a short window on the morning of 21 October picked up the payload; affected teams should assume every credential on the box is gone, rotate all of it, and sweep site-packages for the three injected files. aikido.dev

CMA investigation into Meta — the UK Competition and Markets Authority said on Friday it is scrutinising whether Meta's contractual terms for publishers and advertisers amounted to exploitation, specifically in relation to content used to build its AI services. Data-terms regulation has moved from consultation papers to open cases against named vendors; if you publish, advertise, or buy AI services trained on scraped content, this is the precedent to watch. The probe was announced with no findings published yet — it is an investigation opened, not a conclusion reached. techmeme.com

OpenAI hosted code review — OpenAI's hosted review products for customers' private GitHub repositories — the automated code and security reviews — were failing at unusually high rates, and the status note concedes its availability numbers are blended across every plan, model and error type. If you buy review-as-a-service, the SLA you assume and the number on the page may not be the same thing: a customer on the affected tier experiences worse availability than the published aggregate. Nothing in the status note breaks out a per-tier or per-customer guarantee, which is worth raising in your next vendor review. status.openai.com

Claude Code team's review process — Anthropic has published how its own team reviews a codebase Claude now writes roughly 90% of: a classifier, not the author, decides each change's risk tier, the lowest tier lands with no human sign-off, and the stated target is most changes needing one reviewer at most — who can be the person who prompted the change. The most detailed blueprint yet for near-unattended merging at scale, and unusually honest about where it fails. One CLI pull request burned through more than fifty rounds of overnight agent review before a human halted it; on Anthropic's own repo, a model-written answer misdescribed how issues are managed, and a developer counted over 6,000 issues with reproduction steps auto-closed as stale. infoq.com

“Nobody on the team keeps a complete mental model of Claude Code in their head, and we don't think anyone needs to.”

Scion — Anthropic has written up the system it built after Git stopped coping: daily merged changes in its monorepo climbed from about 2,500 in May to over 30,000 in September, with CI performing around 25 million checkouts a day, so it moved repository state and history onto its own servers behind layered caches. The signal for everyone else is the scale, not the system — Anthropic itself says don't clone Scion, borrow the ideas. Company-reported results: cache hit rates above 99% per tier and an infrastructure footprint 200 times smaller than the old Git mirrors — not independently measured. infoq.com

tsrs — Cloudflare and VoidZero announced a Rust rewrite of the TypeScript compiler that one engineer produced by directing AI agents rather than writing code, claiming type-checking up to 8x faster and about 70% less memory; an independent measurement against a newer baseline put the gain at 2.7x. The board-level point is the clock, not the compiler: work that took Microsoft a funded team two years was approximated in weeks by one person with agents. The release is explicitly experimental — no watch mode yet, and it exits with a different error code than tsc on type errors, enough to quietly break a CI script. infoq.com

AGNTCon + MCPCon 2026 — The Agentic AI Foundation grew from three to ten projects in its first year — Alibaba's agent sandbox, OpenAI's response-format specification, Anthropic's Agent Skills and Hugging Face's agent environments — and the conference program tilted from new protocol features toward security, identity and audit. Agent identity and attestation are becoming standardisation work rather than vendor promises, and that eventually lands in procurement requirements. Claude Code gained support for long-running MCP tasks, so a tool call can now stay open across a whole working session instead of dying after a few seconds; cumulative SDK downloads reached 4.17 billion, up from about 3 billion a month earlier. infoq.com

Artificial Analysis Intelligence Index v5 — Artificial Analysis rebuilt its closely watched index with 60% of the scoring now on private test data, and a day later Google's unreleased Gemini 4 Carbon took the top spot from Claude Opus 5.5; the best open-weight models slid to 12 points behind. Skip the leaderboard and keep the price spread — leading models vary more than 13x in cost per task ($0.96 to $13.04), so model routing is now real budget, and open weights just got weaker as a lock-in hedge. The index says its shift to private tests removes scores optimised against public question banks, not that open models got worse. artificialanalysis.ai

Gemini 4 Argon — Google described an 11-hour unattended run of Gemini 4 Argon over its own 200,000-line Agent Development Kit, in which the agent proposed six changes and developers accepted two, cutting session-preparation overhead about 40% and cold import time 21%. A useful calibration of agent economics: the wins were real, but a third of the output was thrown out on review — review capacity, not code production, is the constraint to staff for. The agent ran all 16,892 open-source unit tests before finishing, and the accepted change shortened a 158-turn benchmark run from 65.6 to 56.1 seconds. infoq.com