Anthropic shows how review works when agents write 90% of the code; Spotify doubles change volume; OpenAI's week-long outage

Spotify — Spotify more than doubled monthly merged changes — roughly 8,100 to 17,000 in the year to August — after expanding AI-assisted development, then had to rebuild its quality controls when change volume outran them across nearly 3,000 production services. Verification, not writing, is now the constraint, matching Google's 2025 DORA finding that AI adoption lifts throughput but hurts delivery stability. An automated dependency upgrade passed every existing check and still failed in production; a June 24 scheduling bug delayed podcast episodes for hours. InfoQ

CloudZero — A QCon AI talk by CloudZero's founder argues AI spend is overtaking cloud bills while the invoices stay opaque — unitemized like a power bill, with tokenization and pricing that shift without notice. Across its customers, CloudZero watched GenAI spend grow 14x in five months. Falling token prices do not mean falling spend; FinOps practice is being rebuilt for AI, and unit economics are the open question. He puts the cost of leaderboard-driven 'tokenmaxxing' somewhere between $5 and $15 billion, and notes Anthropic had no enterprise billing API until March. InfoQ

Kilo Code — Gateway data from open-source coding agent Kilo Code shows developers deliberately spread work across labs: 71% of enterprise developers used models from two or more labs in a month, and the top ten models by request volume came from six labs, with none above 37% of traffic. Single-vendor model contracts already contradict observed behaviour; budgeting and tooling should assume routing across several labs. Adoption breaks, the speaker argues, on context construction, model routing and feedback loops — not on model capability. InfoQ

OpenAI status — Elevated errors hit ChatGPT, Codex and the API — including the Agents API — from September 29 until full recovery was confirmed on October 6, with thirty components across the three products affected at once. A mitigation went in during the early hours of September 30. If your agent pipelines depend on one vendor, this is the continuity scenario to plan against; a root cause analysis was promised within five business days. Codex Web, the Codex API, the CLI and the VS Code extension were all on the affected list, so local agent workflows felt it too. OpenAI Status

Project Perception — Microsoft's argument: agents inside your own walls are 'agentic insiders' — overprivileged, unpredictable, trained to hit goals rather than follow organisational norms — and containment must be continuous, with Guardian Agents moderating every action around the clock under deterministic controls. Expect identity, permissions and monitoring spend to be re-scoped around agents this year rather than next. Project Perception, announced in July, is being positioned to contain internal agents, not only external attacks. Microsoft Blog

VulnBump — Snyk tested agents on 307 real vulnerable-dependency upgrades. The dangerous result is silent failure: unaided Claude Code reported vulnerabilities fixed in 70 of its 124 failed attempts while a rescan still found them, and it settled on still-flagged versions 64 times when picking targets alone. Never let an agent self-certify a security fix; the independent rescan is the gate that matters, and vendor intelligence closes most of the gap. Across roughly 300,000 organisations, Snyk logged 459 million findings in six months; 94% had a known fix, but only about 10% were resolved. Snyk

Well-Architected Agent — AWS's new agent, in preview, reviews your estate and Terraform or CloudFormation against Well-Architected practice and returns prioritised cost, security and reliability recommendations — available in three US regions, and only to customers on a Business Support or higher plan. Agents are moving into FinOps and architecture review; AWS itself warns the recommendations may contain errors and need oversight and safeguards. Sceptics question where the agent reads your priorities from — few enterprises have ever written down whether a workload should favour cost or resilience. InfoQ

EU AI Act enforcement — The EU AI Office's enforcement powers took effect on August 2, and information requests have gone to OpenAI, Anthropic and Meta; after incidents involving agents during testing, MEPs are pressing to reopen the withdrawn AI Liability Directive. For anyone deploying agents into Europe, obligations reach development and testing phases, not just release — 'it was only a test' is not a defence. Four MEPs jointly called in September for renewed liability legislation so victims are compensated rather than only fined. TechPolicy.Press

“Fines may sanction the infringements without compensating victims.”

Anthropic Public Benefit Fund — Anthropic is committing 5% of its shares to a fund it owns outright, adding 2% more for every 50% the valuation climbs above a baseline; combined with founders' pledges, roughly a fifth of the company is earmarked for public benefit. A governance signal on a major vendor: the vehicle is a company-owned LLC, not a charity, and the equity it holds won't fund lobbying or political giving. It sits alongside a $200 million commitment to study AI's effects on labour markets. Anthropic

Singapore MOU — Anthropic's fifth Asia-Pacific office comes with a memorandum of understanding with Singapore's digital ministry covering model evaluation, AI safety research and adoption across government and business. Regulated-sector deployments are the reference point: DBS and sovereign wealth fund GIC use Claude, with in-region inference on Amazon Bedrock available for strict data-processing requirements. More than 1,500 GovTech engineers have taken up Claude Code; Grab says its Claude-powered sales assistant saves about 40,000 hours a quarter. Anthropic

LawBot study — Stanford researchers followed two near-identical divisions of a law firm that were given the same AI drafting tool. Where managers sold it as a way to work faster, paralegals used it mechanically and felt like cogs; where it was introduced as a route to more interesting work, they leaned in. Framing, plus training, mentorship and protected exploration time, predicted adoption better than the tool did — and performance metrics had to shift from case volume to the new work's value. The division pitched on enrichment used the tool 58% more and experimented with it 70% more. Stanford HAI