The Digest

This week in AI and agents

Week 43 · October 19–25, 2026. The stories I would send a client or a colleague, each with a short summary and one line on why it matters if you build with this stuff.

New edition every week. Summaries are my own; follow the links for the primary sources.

  1. 1.

    Scaling online storage to serve over 1 billion ChatGPT users (Part II)

    OpenAI · · Infrastructure

    Part two of OpenAI's Habitat series covers the storage layer under its products: more than 500 petabytes on Azure Cosmos DB, with the largest account serving over 20 million queries per second. The team went from multi-tenant containers to single-tenant and back again, after hot-key throttling, request-tier quality of service, and an improved serverless mode fixed the noisy-neighbor problems that pushed them away the first time.

    Why it matters: A rare public look at multi-tenancy tradeoffs at a scale where a managed service stops being hands-off.

  2. 2.

    Building agents: when to fold subagents into your main agent loop

    Anthropic · · Engineering

    Anthropic's Applied AI team reports that, across several enterprise deployments, a single agent loop with progressively disclosed skills and tools often beat multi-subagent designs on quality, at lower cost and latency. Subagents still earn their place for parallel work, fresh-context verification, and permission boundaries. The post walks through a staged migration and links a migration kit.

    Why it matters: If your agent architecture mirrors your org chart, this is the case for simplifying it before the next model release makes the scaffolding obsolete.

  3. 3.

    The full Claude Platform, available on Google Cloud

    Anthropic · · Platforms

    Claude Models on Google Cloud (Anthropic managed) is generally available: the full Claude Platform, including Managed Agents and computer use, with new features and betas shipping the same day as on the native platform. Access runs through your Google Cloud project, billing goes on a single Google Cloud invoice, and usage counts against existing commitments. Opus 5.5 Fast Mode launches with it, at up to 2.5x faster output.

    Why it matters: Cloud procurement is now a real axis of model choice. OpenAI announced the same move the day before; see below.

  4. 4.

    Decisions API is generally available

    OpenAI · · APIs

    OpenAI's Decisions API answers bounded questions from text, images, or JSON: a probability, a choice from your list, or a score against a rubric, up to 400 questions per request. Built on GPT-6 Luna, OpenAI claims up to 10x lower latency than the same calls through the Responses API. Pricing is input-only at $0.10 per million input tokens, with no output charges, and it now plugs into GPT-Live for voice apps.

    Why it matters: Routing, classification, and gating are where agent stacks quietly burn latency and budget. A purpose-built endpoint changes that math.

  5. 5.

    Introducing Claude Duet, Anthropic's proactive agent app

    Anthropic · · Agents

    Anthropic opened a waitlisted research preview of Claude Duet, a standalone app on iOS, Android, and the web where Claude works as a proactive agent for Pro and Max subscribers in the US. Duet sends a morning briefing, watches your connected apps across more than 6,000 connectors, and keeps working in the background, asking for approval before it sends messages or pays for anything. It runs on Claude Opus 5.5 and shares your existing plan limits.

    Why it matters: Proactive, approval-gated agents are now a shipping product, not a demo. Expect clients to ask for the same pattern inside their own tools.

  6. 6.

    OpenAI models are coming to Google Cloud

    OpenAI · · Platforms

    OpenAI's frontier models and Codex are coming to Google Cloud's Model Garden, purchasable through the Google Cloud Marketplace and drawable against existing GCP spend commitments. The models will also show up in Google Antigravity and the Gemini Agent. General availability is planned in the coming weeks, and Microsoft remains OpenAI's primary cloud partner.

    Why it matters: Both frontier labs landed on Google Cloud in the same week. Model distribution is going cloud-agnostic faster than most roadmaps assumed.

  7. 7.

    Patch the Planet: agents found critical PHP vulnerabilities

    OpenAI · · Security

    OpenAI says a researcher using GPT-5.6 Sol found critical vulnerabilities in PHP, with Wiz independently finding the same flaws, and that frontier AI found them before attackers did. The post lays out the five-stage coordinated response OpenAI ran with PHP maintainers, Wiz, Cloudflare, and CISA, plus $1 billion in subsidized Daybreak access for under-resourced defenders. A PHP version checker ships as a Codex skill so operators can check their own exposure.

    Why it matters: AI vulnerability discovery is now an operations problem. If your stack runs PHP, patch first and read the framework later.

  8. 8.

    Thousands of AI agents uncover new leads in biological data

    OpenAI · · Research

    OpenAI, PerturbAI, and Scripps Research pointed thousands of GPT-Rosalind agents at a whole-brain CRISPR atlas built from millions of mouse brain cells. In 32 minutes the agents produced a reasoning atlas linking gene disruptions to the published literature. At the largest workload they cited more than 700 references where a single agent cited 11, and blinded experts preferred the distributed reports 75% of the time.

    Why it matters: A clean, measured case that splitting work across many agents beats one big context when the job is wide rather than deep.

  9. 9.

    Gemini for government services

    Google Cloud Blog · · Enterprise

    At the Google Public Sector Summit, Google Cloud launched Gemini for government services in preview: the Gemini agent packaged for federal, state, and local agencies, with the governance and security controls public sector work requires. The launch also brings enterprise search to Google Distributed Cloud air-gapped deployments, so sensitive records can be queried and summarized without touching the public internet.

    Why it matters: Agents are entering regulated, even air-gapped, environments. Governance packaging is becoming the product.

  10. 10.

    Four Gemini agents beat Minecraft in 27 hours

    Tom's Hardware (reporting a Google DeepMind paper) · · Research

    A Google DeepMind paper reports four Gemini 4 Argon agents finishing Minecraft end to end with no human input during the run: 27 hours, 2,783 programs written by the agents themselves, 605 deaths on Easy difficulty, and about $520 in tokens. Each agent drove a bot through the open-source Mineflayer library, and the team coordinated through in-game chat and shared notes.

    Why it matters: Long-horizon multi-agent work now has a price tag and a stopwatch: $520 and 27 hours for a task that needs planning, coding, and teamwork.

  11. 11.

    Advancing coding in low-resource programming languages

    OpenAI · · Engineering

    OpenAI published two evaluations: MultiLangBench, which asks a model to build a SQLite-compatible database engine from scratch in 20 languages, and a COBOL migration test that demands byte-for-byte behavior parity. GPT-6 Astra averages 87.4% on MultiLangBench, with COBOL the hardest language at 72.1%, and GPT-6.1 Sol reaches 97.9% on COBOL migration at high reasoning effort. The evaluation suite is public on GitHub.

    Why it matters: COBOL still runs most ATM and card traffic. Legacy modernization just got a credible benchmark and a public yardstick.

  12. 12.

    GitHub Copilot models deprecated on October 19, 2026

    GitHub Changelog · · Developer Tools

    GitHub retired six Copilot models across chat, inline edits, ask and agent modes, and code completions: Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, and Grok 4.5. Each has a named replacement (GPT-6.1 Sol, GPT-6 Luna, Gemini 3.8 Flash, Grok 4.7), but enterprise administrators may need to enable the new models in their Copilot model policies before users see them.

    Why it matters: Model churn is now monthly operational work: pin versions where you can, watch the changelogs, and re-test after every forced swap.

Want this in your inbox? Use the signup in the footer. Want the deeper version? Read the technical articles or the AI strategy notes.