DoorDash counts its agent orders; Jira makes agents measurable; GitHub bundles Copilot into Enterprise

Gemini 4 Carbon — Google's newest top model, built for long-horizon agentic engineering — the company says its internal teams have watched it grind for days through multi-step problems like root-causing a bug across a big codebase. Access starts with vetted cyber-defense teams in the Fairwind Program; general release is promised only "as soon as possible." So the strongest candidate model for hard agent work is not in your stack this quarter — don't write it into near-term plans. Google

Genie 4 — DeepMind's world simulator: describe a scene and it renders a live, explorable 3D world with native sound and characters you can steer mid-stream in plain language. It is a research preview for trusted testers only, with wider access later for Google AI Ultra subscribers. This is the demo-not-shipped frontier: simulation as a product medium, worth watching for prototyping environments and agent training. Sessions run about two minutes before visual detail drifts, and everything it generates carries a SynthID watermark. Google DeepMind

GitHub Copilot — At Universe, GitHub recast itself as an end-to-end delivery system for agent work: Fusion picks the model per task, Copilot projects run agent sessions across many repos in parallel, one session follows you from phone to IDE to web, and automations learn from their own outcomes. The pricing is the PM-relevant part: Enterprise customers now get every Copilot feature with no separate subscription, paying only for the AI credits they use. To carry the load, GitHub says it added 4 million CPU cores this year. The GitHub Blog

“I think all the software we've written to date will amount to less than 1% of the software we'll write in the next five years.”

Jira Autumn release — Agents in Jira now get their own identity on boards and work items under Atlassian's AMP protocol, one settings page to manage them, and — the headline for PMs — a default-on dashboard in the Reports tab that connects agent activity to cycle time, throughput and shipped work. Agent contribution stops being anecdote: you can answer 'did the agents speed us up?' from the same tool you track sprints in. Atlassian

Trello — The rebuilt Trello is generally available: an Inbox that pulls to-dos from Slack, Teams, email and iOS voice, AI Quick Capture that turns a forwarded message into a card with due dates and action items extracted, and a Planner that blocks focus time against your calendar. A signal on where AI lands in mainstream tools: not a chatbot, but capture and triage — the least glamorous, highest-frequency jobs. The rollout is gradual over the coming weeks, hitting Free, Standard and Premium first, Enterprise later. Atlassian

Zest concierge — A restaurant startup launched a concierge that lives in your text messages: it recommends places based on where your credit card says you actually eat (via Plaid), checks availability and books through OpenTable, Resy or SevenRooms, and can join a group chat, run a vote and book for the group. A live case of an agent replacing the app as the interface, with bookings and orders as the revenue line. Zest has $1.8M in pre-seed backing from Alexis Ohanian's 776 and Kindred Ventures. TechCrunch

Super-Agent — DeepSeek has published the source code for a coding agent that doesn't wait to be asked: left running, it keeps an eye on team chat channels, notices bug reports and build breakage as they appear, and drafts candidate fixes on its own initiative. Open-source entrants keep compressing what paid coding agents can charge for; differentiation has to come from integration, governance and trust, not the agent loop itself. It ships as a research prototype — a demo to study, not something to put in a delivery pipeline. Techmeme

AI SDLC transformation playbook — Atlassian published a playbook on its own move to an agent-assisted development process, resting on two investments: a context engine (Teamwork Graph) feeding agents code, docs and decisions from 50+ sources, and measurement via DX that pairs tool telemetry with developer feedback. A named-company baseline for what an AI-native SDLC actually delivers, with the internal numbers attached. Atlassian reports cycle time dropping from 11.5 to 6.7 days per issue, agent loops auto-resolving 51% of flagged vulnerabilities, and graph-connected agents testing 44% more accurate on 48% fewer tokens. Atlassian

Cisco's legacy-code bet — Cisco's president gave a 27-year-old engineer he found five levels down the org chart unlimited AI tokens and four weeks to prove a two-person team can modernize the company's legacy code, with a stated goal of refactoring 98% of it inside 24 months. Agent leverage is making 'small team, huge mandate' a credible staffing pattern — and shifting who gets trusted with bet-the-company work. Padmanabhan Krishnamurthy joined via Cisco's acquisition of his startup EzDubs. Fortune

Claude Code extension backdoor — A tampered build of Anthropic's official VS Code extension sat in Cursor's marketplace for almost twelve hours, installing a persistent SSH backdoor and harvesting host details from developer machines before it was pulled. Your coding agents' distribution channels are supply chain: worth a vendor review and an incident plan before, not after, this reaches your team. At most 5,000 users were hit; anyone who installed or updated between 5:27 am and 4:35 pm ET on Thursday should assume affected, and Anthropic has not publicly warned customers. Ars Technica

Spotify AI remixes — The Financial Times reports, citing sources, that Spotify plans to let subscribers create their own remixes and covers for about $5 a month, targeted for early 2027 — unconfirmed, and contingent on licensing deals closing. A pricing signal for consumer AI add-ons: a flat monthly fee attached to a generative feature, not bundled into the base subscription. The report is by Anna Nicolaou; Spotify has announced nothing. Techmeme

__usage_ai-training — OpenAI, Google, Microsoft and Anthropic have committed to honouring one shared robots.txt token — __usage_ai-training — that lets a site refuse AI training across all of them with a single line, instead of a rule per crawler. If your product publishes web content, one robots.txt entry will soon carry a preference most big AI firms read; it's voluntary, though, and untested in court. Anthropic already honours it, Microsoft starts 1 February 2027, Google in 2027; Microsoft also lets sites control, page by page, whether content may shape answers in Copilot and Microsoft 365. The Verge

Anthropic S-1 — Anthropic filed its IPO paperwork, aiming for a Nasdaq listing under the ticker ANT, with reports suggesting the raise could reach $100 billion at a $1.5 trillion valuation. The company behind Claude going public means its pricing and roadmap commitments will soon face quarterly scrutiny — relevant to anyone with multi-year vendor bets on Claude. Anthropic

OpenAI funding round — OpenAI cut its target from $30 billion to $20 billion at a $1.25 trillion valuation after lukewarm investor interest, having secured about $11 billion so far, per the FT and WSJ. The funding wobble at the sector's flagship is your warning that agent-feature business cases will get harder scrutiny next year. Trending Topics

Claude vulnerability research — Anthropic says an internal research model chained real zero-days into working exploits against Chrome, Safari, Android, Windows and Linux, all responsibly disclosed. Breach expertise that once lived in a handful of specialist teams is now a model capability — relevant to your security roadmap and your vendor-risk conversations alike. 62 CVEs were assigned in Chrome alone; one Linux chain went from network packet to root shell in under an hour. Anthropic

GPT-6.1 Terra — OpenAI's new mid-tier model handles routine computer use — reading screens and clicking through high-volume office tasks — at $1 per million input and $5 per million output tokens, half its predecessor's price. The unit cost of commodity computer-use keeps halving; that is the line your agent-feature economics ride on. The Next Web