Anthropic's own AI bill tops payroll; Devin gets org-chart budgets; Grok caps bite

🔥The one thing to know

🧭How teams are doing it

• Atlassian's compensation-planning team — A five-engineer team inside Atlassian shipped a production compensation-planning app with agents doing the coding, then measured it against a comparable product built the old way: roughly five times the output per engineer, and nearly five times as many Jira tickets per person. Writing more, better-scoped tickets was the control that kept parallel agents from inventing requirements. For anyone running several agents at once, this is evidence the manager's leverage moves upstream into scoping and acceptance criteria rather than into review volume. A source-blinded judge scored 300 tickets at 4.47 of 5 for writing quality against 2.72 for traditional ones, and rated 83% 'agent-ready' versus 6%. Inside Atlassian

• Atlassian's 2026 State of Developer Experience — Atlassian's annual developer survey finds 87% of developers say AI helps them ship faster, yet 72% report the same or more friction than before AI arrived, and 65% now spend regular time correcting model output. The share saving ten-plus hours a week has collapsed from 68% last year to 25%. The useful split for leaders: teams at organisations with a clearly communicated AI strategy are twice as likely to feel less friction, so the strategy — not the tooling — is the lever to pull. Inside Atlassian

🛠Agents and plans

• Cognition Devin Budgets — Cognition is giving enterprises spending budgets for its Devin agent that mirror the org chart: one central pool split as deep as three levels across departments, teams and projects, with engineers warned and able to request more capacity before a session is halted. Admins can set a rule that lifts an engineer's cap when their Efficiency Score is strong. This is agent FinOps landing inside the product itself — worth knowing before your finance partner asks who owns AI spend per team. A planned soft-limit mode would step a long-running job down to a cheaper model rather than stop it outright. The New Stack

• Grok Bot — Users burned through weekly allowances far faster than usual over Halloween weekend, and Musk confirmed on X the caps were deliberate, not a glitch: serving the agent costs more than it brings in. The product has roughly 800,000 paid weekly users and no standalone price — it rides along with $30-a-month subscription tiers. The first plain-language admission that a flat-rate agent seat can be underwater at scale; expect other vendors to tighten caps quietly rather than raise list prices. Musk gave no fix, no price change and no reset; some users hit 100% of their allowance with four days left before the weekly reset. TechCrunch

• Claude Code — Anthropic has confirmed new Claude Code limits meant to steer the agent away from recreating a specific existing game, on the grounds that copying someone's IP breaches its usage policy. Users are also reporting that reverse-engineering work is being blocked, and complaints are piling up online. A vendor deciding mid-task what counts as infringement is a new failure mode for teams doing legitimate interoperability or clean-room work — worth a line in your agent-use policy on what to escalate and where to route disputes. Anthropic has not published where the boundary it is enforcing actually sits. Tom's Hardware via Techmeme

🧰Tools around the agents

• Momentic — Momentic's agent Mo — now generally available, out of beta — runs end-to-end tests written in plain English and stored as YAML files in your repo, finds controls again when the UI shifts, and runs in CI as a required check. Your coding agent can author a test for its own change via MCP and save it beside the code. Verification is where most agent-heavy teams feel the squeeze now; this is one answer that keeps the tests reviewable in the pull request. Quora reports daily test execution fell from seven hours to 30 minutes after adopting it. Momentic

💬What people are saying

• Hacker News on San Francisco's jobless tech workers — A thread on laid-off Bay Area engineers missing the AI boom fights over causes: the argument carrying the thread blames the industry sinking its cash into AI data centres rather than immigration or offshoring, while others counter that most of the region never worked in tech and sympathy is misplaced. The hiring signal for you: senior engineers displaced by AI are not being absorbed by AI companies, and commenters report startups still hiring remotely even as local candidates stall — your hiring bar and your team's job-security anxieties are both live issues this week. One thread commenter claims most of 2026's software is now written, reviewed and tested by AI, whether or not practitioners like the practice. bloomberg.com · from the Hacker News discussion

🌍The big AI stories today

• Anthropic S-1/A — The amended filing shows Q3 revenue near $17bn — twelve times a year earlier — against a $23.3bn paper loss and $895bn of compute commitments, ahead of an IPO before Thanksgiving. It is the economics of the vendor behind your Claude seats on one page. SEC

• NYT v. Brave — The New York Times and three other publishers sued browser maker Brave for allegedly selling scraped, paywalled content to AI companies including Anthropic and Perplexity — the sourcing behind agent answers is back in court. The Verge

• SUSE layoffs — Linux vendor SUSE is cutting about 10% of its employees, telling staff that automation and AI are reshaping how software companies operate — the AI-jobs conversation has reached infrastructure vendors. The Register via Techmeme

  “automation and AI are fundamentally shifting how software companies operate”

• WorkOS breach — Auth provider WorkOS says an autonomous AI penetration-testing framework called ARTEX reached its Salesforce and Snowflake data on October 30, with production customer systems untouched — rotate keys if you are a customer. Hacker News

💡One takeaway