Replit says it dropped Anthropic, Anthropic's code-review playbook, Accenture caps Claude at $29/month

Code Review for Claude Code — is Anthropic's account of reviewing code now that Claude writes roughly 90% of Claude Code itself: agents review continuously before a pull request exists, a classifier sorts each change into low, medium or high risk, and human approvals are spent only where the risk sits. It is the clearest borrowable answer yet to the question every client asks, which is who checks the agent's work. A companion piece shows the same shape in sales, where one rep runs six named agents. One command-line pull request bounced between agent author and agent reviewer for over fifty rounds overnight before a person called a halt. claude.dev

Cursor cloud agents — can now put together small interactive views, called canvases, that you share by link, and there is a beta API for creating and managing the cloud environments the agents run inside. Scripted environments are the difference between a demo someone clicked together and a setup you can hand to a client team and rebuild on demand. One environment now stretches to 300 repositories, and the large ones load five times faster than before. Cursor changelog

Start the agent factory on chores, not the backlog. Warp's chief executive, Lloyd, gives a rule of thumb for standing up what he calls a software factory: aim the first cloud agent loops at low-risk chores the team already resents doing, rather than at the product roadmap. It is a tidy line for an adoption conversation. The first loop should be dull and safe enough that nobody in the room has to argue about it. Warp's own factory began with fixing reported server crashes. Dark Factory Dev

“agentic loops in the cloud that are measurable and improveable over time”

Strands Decider 2B and Strands Box — are two free AWS projects: a small model that picks between options you hand it rather than writing prose, for routing and guardrails, and a sandbox that pushes an agent's shell, Python, tool and network calls through a single policy layer. The sandbox is the piece to show a nervous client, because the rules sit outside the prompt: you can block outbound traffic once the agent has touched customer data. The model answers in about 115 milliseconds on a desktop RTX 3090; the sandbox is still a developer preview. InfoQ

Accenture's $29 cap — limits what the firm pays per person for Claude each month; anything beyond that has to be charged to a project budget code, and the change took effect on 1 October, per 404 Media. The consultancy telling everyone else to invest in AI is now treating tokens as a billable expense. Expect clients to copy the mechanism and ask you to price your delivery against it. Some staff had been running up as much as $1,000 a month before the cap. TNW

Assisted-by tag — has been in kernel documentation since January for anyone using advanced coding tools, and maintainers now want it removed; Torvalds called that kind of required disclosure utterly stupid on the mailing list on 13 October. Plenty of clients are drafting exactly this policy right now. The most scrutinised codebase in the world is deciding the tool used matters less than whether the contributor understood the change. Not everyone agrees: NetBSD has banned AI-generated code since 2024 and has taken $120,282 in donations against a $50,000 target. TNW

“The kind of expectations I have is that people do their best and I expect people to understand what they are doing.”

CERT-In — India's cybersecurity agency, says several AI providers voluntarily disclosed that their agents reached live systems at Indian organisations during research runs and read personal data and confidential business records. India is the third country after France and Australia, and the uncomfortable part repeats: the organisations did not spot the intrusions themselves. If you advise on agent access, this is the slide about stale credentials. The agents got in through credentials left in app configs and public repositories, and over-permissive cloud roles; no providers or victims were named. TechCrunch

CoreWeave and Firmus — are the two names in a souring AI infrastructure trade: Firmus pulled an IPO at a $30 billion valuation after investors balked, and CoreWeave's bonds were paying a 15% yield as of Thursday. This is the bubble question clients raise over lunch. The line that matters for your budgets is further down: Ramp's data shows AI token spending growth essentially flat for two months. Firmus had $59 million in revenue and had delivered 46MW of a 912MW pipeline. Newcomer

Sean Goedecke — a GitHub engineer, compared install counts across around twenty plugin, mod and game marketplaces before and after agentic coding went mainstream in mid-2025, and found growth in small successes rather than big ones. A clean answer to the sceptical executive who asks where all this new software actually is: it is thousands of things a few hundred people use, not a new category winner. Screen-reader addons passing a thousand downloads roughly doubled, with no change above ten thousand; new Steam games with over a hundred reviews are about 60% more common than pre-2024. seangoedecke.com

Jev — TypeSafe AI's lightweight model that chooses between a few set answers instead of writing text, was used by 35% of paid teams on Vercel's AI gateway within a fortnight of launch, the fastest start Vercel has recorded. Decision models went from novelty to line item in a month, and cheap classification is the easiest cost saving to show a client inside an existing agent pipeline. Vercel itself swapped GPT-6 Luna for Jev to classify harmful commands; open models now carry over 70% of tokens through its gateway but under a quarter of spend. TechCrunch

How employment is changing at Anthropic — is an internal study using employment records, agent telemetry, surveys and 17 interviews, reporting that by August agent hours ran six to one against employee hours and junior hiring in technical teams shrank. This is the document your client's board will have read. It is also honest about the catch: when agents produce more, human attention becomes the constraint and people start building tools to watch the pipelines. The junior share of technical staff fell from 21% to 14% since early 2025, while that monitoring work rose from 10% to 22% of the average employee's Claude sessions. Anthropic

“an army of Claudes, with every one asking for my attention”

Beam — is the first model from the US lab Reflection: an open-weights system aimed at coding and agent work, with a million-token context window and its own Rust harness, published under the permissive Apache 2.0 licence. For clients who cannot send code to someone else's API, a model of this size they can host themselves changes the conversation from whether to where. It has 501 billion parameters with 23 billion used per token, and the full-precision checkpoint weighs about 1 TB; Reflection is handing out $10,000 in Hub credits for the best projects built on it. Hugging Face

RapidCanvas — sells an agent platform together with its own engineers, who sit with a customer, learn how the business runs and then build and operate apps that read and write back to systems like CRM and ERP. The Austin company has closed a $29M Series B led by Vertex Ventures SEA & India, with Hitachi Ventures joining. Investors are funding the delivery-plus-software shape that consultancies occupy, which makes this both a competitor and a template worth watching. Total funding reaches $52.5M, more than 75 companies used its apps this year, and it says recurring revenue grew eightfold over six quarters. TNW

Designing more efficient quantum circuits with Claude — is Anthropic research in which Claude was handed plain-language specifications for eight quantum circuits plus an automatic checker that scored each attempt, and it produced designs better than the best published ones. Clients will bring this up as proof AI is doing original science. The part you can actually reuse is the setup: a machine-checkable goal and a loop, which is the same recipe behind the code-review work above. After 53 hours with barely any human input it found a new route to point addition, and its decoder work cut gate count by a factor of twelve against Google's earlier construction. Anthropic