🧩 Agentic Engineering for Engineering Managers

Atlassian Rovo Dev study — Atlassian compared repos that adopted its Rovo Dev agent with statistically matched repos that didn't, across 3,400 repos from 2,500 customers. Adopters merged 19% more pull requests per month, with quiet repos gaining most (37–51%). A survey of more than 6,200 developers put the time saved at 2–3 hours per week. It's the rare AI productivity claim with a comparison group — a template for measuring your own rollout instead of pointing at usage dashboards. Where 3–5 people in one repo adopted, throughput rose 59–87%: the lift is a team effect, not an individual one. Atlassian (Inside Atlassian blog)

Atlassian DevEx survey — Atlassian surveyed 3,500 developers and managers across six countries: 99% now save time with AI and 68% save more than 10 hours a week — yet half also lose 10+ hours weekly to non-coding friction. The share who say leaders don't understand their pain points jumped from 44% to 63%. Banking the AI hours while ignoring the friction is exactly what widens that gap; the fix starts with asking where your developers' hours actually go. The biggest reported sinks were hunting for information, getting new tools working, and switching between tools — with tech debt dropping off the worst-five list this year. Atlassian (Bitbucket blog)

PostgreSQL 19 beta — Roughly 8–10 significant features were pulled from PostgreSQL 19 during beta, against a usual one per cycle. A committer's postmortem partly blames LLM-assisted code review for surfacing long tails of edge-case defects on features written years earlier, plus a security-fixing crunch that kept senior reviewers busy from April to August. The effective quality bar under your team's code can rise after it ships; plan review capacity for the long tail, and don't read late reverts as AI slop. One feature drew about forty defect reports, each averaging a three-line fix — the per-item overhead is what ran out the clock. Peter Eisentraut

Agent ownership checklist — A sponsored New Stack walkthrough (by an engineer in Andela's talent network) treats a production agent like any other production service: a named owner who takes the pager, a tool allowlist that denies by default, execution budgets, health checks, and human sign-off before anything risky runs. On-call for agents is the unresolved part of most rollouts; decide who gets paged before the 2 a.m. failure, not after. The core pattern separates a recommendation from its execution, so approval stays a human decision the agent can't take for itself. The New Stack

Claude-built dashboards — At a ~500-person company, three data platform engineers spent eight months of side time on a Claude-generated HTML dashboard tool. It now matches Looker's weekly user count, has cut Looker usage by about a third, and cost 325 commits, roughly 74,000 lines of code and $6.4k in tokens. When a dashboard costs pennies to build, the $150k-a-year BI licence becomes a negotiation — and the real engineering went into the security model, since every dashboard is unreviewed JavaScript by design. Per dashboard, the analyst agent burns about 1 million tokens, and a model switch dropped the cost from around $2.50 to around $0.50. David Lequin

Claude Duet — Anthropic opened a research preview of an always-on agent that sends a morning brief across your connected apps, accepts tasks by message, and asks approval before sending payments or messages. It runs on Opus 5.5, reaches more than 6,000 connectors, ships on iOS, Android and the web, and competes with Meta's Muse and OpenAI's Dots. Nothing to buy for teams yet: the waitlist covers US Pro and Max subscribers, and Team and Enterprise are not included yet, with no timeline given. Sensitive steps are gated by a dedicated classifier plus a safety prompt that blocks the strictest information tier from being shared. Anthropic

Agent Skills — Anthropic donated its Agent Skills standard — folders that teach an agent a task, claimed to run across 70+ platforms with 1 million skills published — to the Agentic AI Foundation under the Linux Foundation, where MCP already lives. The foundation also took in OpenAI's Open Responses, Hugging Face's OpenEnv and Alibaba's OpenSandbox at its conference this week. The format is now vendor-neutral, safe to build internal skills on; but a skill is still executable code, and the standard ships with no signing or provenance chain. A Snyk audit in February found 76 malicious skills among 3,984 scanned on two public directories. Simon Willison

Claude Team plan — Standard seats cost $25 per member per month ($20 billed yearly); Premium seats cost $125 ($100 yearly), with a two-member minimum and a 150-seat ceiling before Enterprise. Weekly usage limits apply per member, not across the team: Standard carries 1.25x the Pro per-session allowance and Premium 6.25x, and admins can mix seat types so heavy Claude Code users get the bigger allowance. Useful when sizing seats: give Premium to your power agent users rather than raising the whole team's bill. Organisations can prepay usage credits so members keep working past their weekly limit. Claude Help Center

AWS Continuum — AWS moved Continuum, its code-vulnerability agent, from preview to general availability. Point it at a repository and it hunts for real bugs, confirms they're exploitable in your AWS setup, and files pull requests with candidate fixes. Its headline number — 89% of tasks passed on CyberGym-E2E, a public test giving an agent 90 minutes to fix a memory-safety flaw in C or C++ with no hints — comes from a run AWS graded itself. Automated vulnerability fixing is about to be a procurement conversation; know the caveat before anyone repeats the number upstairs. Only 37.8% of its patches also fixed the specific historical bug the benchmark secretly checks, versus a previous public high of 26.2%. The New Stack

Ramp Laminar — Ramp assessed managed agent APIs and third-party platforms, then built its own harness — Laminar — to a first version in six weeks; it now runs every agentic product at the fintech, keeping agent sessions separate from the sandboxes that execute their code. LangChain's CEO called the design part of a broader industry shift. The build-versus-buy question for agent infrastructure is now real, and the measurable savings live in the harness: routing each task to the cheapest capable model. LangChain's own router cut median cost per agent thread from $2.61 to $0.94 across 973 threads, with no significant drop in merged-PR rate. The New Stack

Torsten Slok — Apollo's chief economist spent the week arguing AI has delivered no financial payoff outside tech: revenue per worker and profit margins are climbing for the Mag 7 and flat for the other 493 S&P companies, with no productivity lift visible — and mainstream corporate America starting to treat token spend as a cost centre rather than a growth story. These charts will reach your CFO. Be ready to show where your team's AI spend pays back in delivery terms, not adoption stats. He also argues a unit of compute costs more to make than it earns back for buyers, and that eight in ten of the biggest token spenders are venture-funded — investor money, not revenue, paying the compute bills. Apollo (The Daily Spark)

“AI may be a great calculator, but in business, intelligence is mostly judgment.”

Kimi K3.1 Pro — Moonshot released an open-weight model that nearly matches Claude Opus 5.5 on coding-agent tests and costs $2/$10 per million tokens; weights are due November 4. A cheap near-frontier option if your team routes its own agents. Moonshot AI

ProjectDiscovery research — Researchers planted a backdoor in a small open model for under $50 that stole .env credentials through Codex CLI the moment a trigger phrase appeared. Treat edited weights from public hubs as unvetted code — nothing most teams run today would catch this. ProjectDiscovery

Railway breach — Someone may have read Railway customers' environment variables, some of them decrypted; Railway paused all deployments and says it's part of attacks on platform companies aimed at AI inference providers. If your team stores model API keys or registry credentials there, rotate them now. Railway blog