Agentic Engineering for CTOs - Weekly

🔥The week's one thing

🏢How organisations are changing

• METR's Anthropic audit — was nearly three months inside Anthropic (August 12 to October 29) and concluded the lab is improving model capabilities at roughly twice its historical pace, mostly because its researchers use AI to do the research itself. The catch for buyers: Anthropic redacted a majority of the evidence METR found most important, so the speed-up claims are directional, not independently checkable. Benchmark growth is 1.9x the pre-October-2025 rate, and METR is confident it is not from faster compute scaling. METR

• Conductor — is joining Meta. Cloud workspaces go read-only on November 13 and shut on November 20, when data deletion begins; new sign-ups stopped November 4 and the local app gets fixes only until early 2027. A working agent tool can vanish within two weeks of an acquisition — inventory which of your tools are one deal away from shutdown, and check your exit and data-export terms now. Conductor was founded in 2024, launched its core product last summer, and its agents are Anthropic's and OpenAI's models, not its own. Conductor

• Railway — saw an employee's agent run automated signups and tests against rival cloud Primcloud, whose founder publicly accused him of attempted breach; the employee stopped the harness, and Railway's CEO has now banned non-manual testing on outside companies. Liability law for agent misconduct is still forming, but the reputational hit lands on the employer immediately. The tool ran on DeepSeek and was halted within 15 minutes; Primcloud has no bug bounty and treats any probing as a security violation. Business Insider

  “You operate the machinery, you are responsible for that what that machinery does.”

• Microsoft — is offering buyouts to 60,000 people as spending priorities move toward AI, the New York Times reports — a striking step at a company long known for stability. The largest employer in software is converting headcount budget into AI spend at scale; your own workforce plan should assume others follow. Staff checked an internal site within the hour of the announcement to learn whether they'd been offered a package. New York Times via Techmeme

🧾Vendors, pricing and risk

• Claude Code — was handing users' account email addresses to websites inside the User-Agent header. Anthropic had put the address in the model's context for author details on code; when the model wrote web commands itself, it sometimes reused it as a contact address. Personal data left the building without consent — a live GDPR question for European firms — from a tool your developers run daily. Anthropic has apologised; the address disappears next release, though the stable channel lags behind at 2.1.311 and the bug report is unresolved. GitHub (anthropics/claude-code)

• Ramp's spend data — shows AI spend grew just 1.5% across August and September after 42% growth in July, even as token volumes kept rising — price cuts by OpenAI and Anthropic, plus customers routing work to cheaper models, are capping vendor revenue. It's a buyer's market: renegotiate now, and don't read your vendor's growth story as evidence of your own capability gains. Anthropic says its own price cuts cost it about $600 million of revenue growth; October spend re-accelerated 8.9%. TechCrunch

• GitLab 19.5 — turns external coding agents into a governed part of the platform: Claude Code, Cursor and Codex sessions can be registered into AI Governance with audit events streamable to a SIEM, and the MCP server is generally available. This is the working pattern for admitting agents to the SDLC without losing the audit trail — which agent, which user, which project. MCP rate limits are 60 requests a minute on Free and 600 on paid tiers; a new auto mode skips per-command approval, off by default. InfoQ

📰News the board will ask about

• Apple — has cut some engineers' daily AI spending cap from $300 to $200, with reductions to as little as $100 planned, Bloomberg's Mark Gurman reports citing sources — and staff are reading the cost-cutting as a layoff signal. Even the cash-richest company treats per-engineer AI spend as a managed budget line — worth mirroring before your own bill surprises the CFO. Apple has not confirmed the figures. Bloomberg via Techmeme

• Claude Fable 5.1 — is generally available at $10 per million input tokens and $50 output, with cache reads at $0.25 — roughly 25% cheaper than Fable 5 for typical workloads, and up to about 45% for agentic ones. Pricing and data terms are moving in customers' direction: new Enterprise Frontier Safeguards promise zero-data-retention privacy with data stored in customer-controlled cloud, due to begin later this fall. Mythos 5.1, the same model with different safeguards, is limited to trusted-access cyber and life-science programs. Anthropic

• OpenAI — has closed roughly $11 billion in new funding at a $1.25 trillion valuation, Bloomberg reports citing people familiar with the matter; the round could still grow to as much as $20 billion. This completes the raise OpenAI had already scaled back from $30 billion — capital still flows to model vendors even as customer spend growth softens. OpenAI has not commented officially. Bloomberg via Techmeme

🌍The big AI stories this week

• GPT-6 Cyber — launches with a Critical cybersecurity rating from OpenAI: it reproduced 1,724 Linux kernel vulnerabilities, built 948 working privilege-escalation exploits, and documented 81 new high-severity Firefox flaws that Mozilla is now patching. All figures are OpenAI's own and access is limited to vetted Daybreak Red customers, but offensive capability at this scale in a commercial product changes your threat model. GPT-5.6 Sol managed 166 exploits with the same tooling. SecurityWeek

• Anthropic S-1/A — shows about $17 billion of Q3 revenue — twelve times a year earlier — a $23.29 billion net loss driven mostly by convertible-note accounting, and roughly $895 billion of purchase and lease commitments, substantially all for compute and infrastructure. Your biggest agent vendor has mortgaged its future to data centers; expect that bill to reach your contract terms. The IPO is set for before Thanksgiving, led by Goldman Sachs and Morgan Stanley. Trending Topics

• Wyoming data center — is Google's first in the state, near Cheyenne, powered largely by a 702 MW natural-gas fuel-cell plant with carbon capture and air-cooled to spare local water. Power, not chips, is becoming the constraint on your AI vendors' capacity — this is what securing supply looks like two years out. The fuel cells come from Bloom Energy under a 20-year power deal. Google

• Binance heist — is now officially attributed to North Korea's TraderTraitor group; the entry point was a compromised DevOps engineer's laptop, and most of the stolen funds moved in about 20 minutes. A single engineer laptop with deployment access remains the softest target in any organisation — worth a fresh look at your own. Binance says it has covered about 95% of the stolen amount and resumed first-phase withdrawals on November 1. BleepingComputer

• NYT v. Brave — and other publishers sued Brave in federal court in Manhattan, alleging it built a 30-billion-page index publishers couldn't block, sold access to AI companies including Anthropic, Perplexity and Cohere, and ships a paywall-bypass feature. Content-supply litigation is converging on the AI stack your agents fetch through — the legal ground under web-fed agents is shifting. The Times seeks an injunction and damages. The Verge

💡One takeaway