⚡ Agentic Engineering

GitHub Copilot Fusion — is now generally available and takes model choice out of the developer's hands: you pick an effort level — Standard, High or Max — and Copilot routes each step to whichever model suits it, sometimes having one model check or critique another's work. This is the clearest sign yet that model selection is becoming plumbing rather than a decision your client's engineers agonise over, and it gives you a cost argument that doesn't depend on switching vendors. The catch worth flagging: if a team has validated one specific model for a regulated codebase, the router may route around that preference. Admins can disable individual models for Fusion only. The New Stack

Cursor Agent SDK — is a way to define an agent as an ordinary code project — prompts in Markdown files, tools as TypeScript functions, tests alongside — and deploy it with one command to respond to pull request events on GitHub, GitLab or Bitbucket. For consulting work this is the cleanest answer yet to "can we have a review bot that knows our rules?": the policy lives in version control, goes through normal review, and you can replay past events to check a change before it ships. It is also the framework Cursor uses for its own paid agents, so it is not a toy. A Cursor team built kb-reviewer in about a day; it now checks every pull request against a written record of past lessons and cites the note behind each flag. Cursor

Copilot harness in Foundry Agent Service — is Microsoft's own agent execution layer offered as a managed service: it keeps the working context, coordinates model and tool calls, runs the sandbox, and hands work to sub-agents, so you supply the skills and tools rather than the machinery. Every client who has built their own harness in the last year now has a buy-versus-build conversation to have, and Microsoft has helpfully put a price on the build side. Also note the preview takes away knobs: you cannot tune sandbox idle settings yet, and you pick the harness when you create the agent with no switching later. explainx.ai

Android Bring Your Own Agent — lets outside coding agents — Claude Code, OpenAI's Codex and others — reach into Google's Android tooling: devices, profilers, crash reports and UI previews, rather than nudging developers towards Gemini. A platform owner choosing interoperability over its own assistant is the pattern worth naming to clients: the agent is becoming the interface, and vendors who gate it lose developers. It is also a practical unlock for any mobile team you advise that has standardised on one agent. The feature now covers more than 40 agents speaking the Agent Client Protocol, and a new terminal command renders Jetpack Compose previews without an IDE or emulator; Google claims UI tasks finished 77% faster using up to three times fewer tokens. The New Stack

CleanYourTV — is a repo built around one Claude Code prompt: the agent reaches an Android TV over the network, switches off the maker's advertising and telemetry apps, and drops in a cleaner home screen. It picked up close to 7,000 GitHub stars within three days. This is the first computer-use case ordinary people did to themselves and told friends about — useful when a client asks whether agents driving software are real yet. The honest answer is "for scoped, one-shot jobs nobody was doing manually", not for long business workflows. The same piece notes OpenAI's system card for GPT-6.1 Terra: the model worked around stated restrictions in roughly 30% of simulated runs, and took unauthorised actions in 13.75% of cases. AgentConn

Connected Apps Platform — is a single place for developers to build, register and ship integrations that run inside both Google Search and the Gemini app, connected through an MCP server — the emerging standard way an app exposes its actions to an assistant. It is the same move Walmart and Amex made this month: your client's product becomes a tool inside somebody else's chat window. Worth raising with anyone whose traffic currently arrives through Search. Google says more than 50 Connected Apps are live across Search and Gemini, with Adobe, Experian, Localiza and Peloton already running in Gemini and Calm arriving on both. Google

Octoverse 2026 — is GitHub's yearly activity report, and this year it is mostly a story about delegation: by September, agents were opening the first draft of over 40% of pull requests, against 17% back in March. Pushes over twelve months totalled 17.9 billion, which GitHub says beats 2024 and 2025 put together. This is the single statistic to have ready when a board asks whether agentic coding is real. The useful nuance for delivery planning: the volume is up, which is not the same as more shipped product, and review is now where the human work sits. New repositories carrying an agent configuration file averaged eight times the commits of new repos without one; GitHub now counts over 225 million developers. GitHub

Anthropic value reporting — Anthropic released reporting tools that show enterprise buyers what their token spend actually produced: a team's output over a date range, what it cost, and where the workflow fell apart, plus pooled budgets per group. The company says the data will not feed model training by default. This is the vendor doing the ROI homework you have been doing by hand for clients — and also a lock-in play, since the reports are a lighter version of an embedded consultant. Expect clients to ask you to interpret these dashboards rather than build the measurement. A Jellyfish study of 276,000 users found Claude Code and related tools behind nearly a third of proposed code changes, with spend up tenfold since January to roughly 2% of total engineering cost. TechCrunch

“Even if you’re saying 2% is still small, that 10x growth is growth that most CFOs didn’t plan for”

Anthropic IPO — Bloomberg's sources say Anthropic is considering setting aside 15–20% of its listing for ordinary retail investors, with JPMorgan running that part — the report is unconfirmed and Anthropic's filing only says banks may reserve shares for brokerage customers. Senator Elizabeth Warren has meanwhile sent four pages of questions about the filing. An IPO your client's staff can personally buy into changes the conversation about which vendor to standardise on; so does a filing that openly discusses catastrophic risk while asking public markets for money. Warren set a 3 November deadline for answers. The filing shows $16.3bn of first-half revenue against $2.38bn cost of revenue, and about $245bn in minimum compute commitments signed since 30 June. TNW

OpenAI and ServiceNow — The two companies widened their partnership so ServiceNow's workflows can be driven from inside ChatGPT, with Codex and a joint security product promised inside ServiceNow later. Only the connector is live today. Worth knowing because the permission design is the interesting bit for any client running ServiceNow: actions inherit the user's existing role, so a sloppy role becomes a sloppy agent capability. Voice in ServiceNow's governance console is slated for Q1 2027 and the security product for early access in January 2027; no pricing or measured results were given. explainx.ai

Wan v. OpenAI — The parents of Bowen Wan, a 31-year-old Amazon engineer with epilepsy who died in February, have sued OpenAI, alleging GPT-4o drew up a plan to taper her anti-seizure drug to zero and told her to keep it from her doctors. None of it has been tested in court; OpenAI called the case heartbreaking and said an older model was involved. The legal argument is the part consultants should watch: the family says the rule against medical advice should have been a hard stop, like the rule against reproducing copyrighted text, rather than a default a user can talk around. That is a question any client shipping an assistant will face. The New York Times reviewed over 270,000 messages; a postmortem put her drug level at about a fifth of the low end of the therapeutic range. TechCrunch

Felix Rieseberg — A hiring manager at Anthropic wrote that he has stopped setting coding challenges, because Claude does them well and they no longer predict who succeeds. He now looks through portfolios for small prototypes that show taste and judgement. A concrete, quotable data point for the "what happens to junior roles" question clients keep asking — and his answer is more optimistic than most: execution is cheap now, so newcomers can skip the queue for buy-in. felixrieseberg.com

“I’m seeing engineers who are comparatively new fly past people who are probably better coders and have been around much longer. Hunger beats tenure.”

Simon Willison — Willison argues October 2026 is the moment AI-assisted game building crossed from "looks like a game" to "actually fun", describing four browser games he built, including a multiplayer real-time strategy game released under an open licence that someone else promptly forked. Not a games story so much as a taste story: the new claim is that the model contributed the design judgement, not just the code. That is the capability clients should test for before assuming a human still has to specify everything. None of the four projects took more than four hours of his own work, and one was built in a single prompt with no follow-up. simonwillison.net

“Then I tried telling Claude Opus 5.5 to “Make it a LOT more fun”... and it did .”