Agents crossed into production, and trust paid the bill

GitHub put numbers on the agent shift on 8 Nov: more than 40% of pull requests are now initiated by agents, with 225 million developers and 34.3 billion commits on the platform. That followed its 1 Nov disclosure of a sixfold increase in monthly commits to 7.4 billion, requiring an additional 4 million CPU cores and 160 petabytes of storage.

The product line followed the metric. On 7 Nov GitHub added chat and VS Code agent usage to the Copilot impact dashboard. On 5 Nov it released gh-aw 1.2.0 with security enhancements, improved OpenTelemetry support and model routing changes. On 7 Nov gh-aw 1.2.1 added durable workflow continuation and a specification-aware PR reviewer.

This is the week agentic coding stopped looking like a feature layer and started looking like an infrastructure load. On 2 Nov, GitHub COO Kyle Daigle said AI agents could become bigger consumers of GitHub services than traditional developers. The important part is not the phrasing. It is that the same fortnight contained usage dashboards, workflow durability, model routing and capacity disclosure.

Cursor had the hardest trust week. On 31 Oct it reported that Claude Code extension version 2.1.313 in its marketplace had been tampered with and installed a persistent SSH backdoor on affected devices, with up to 5,000 users potentially impacted. On 2 Nov, reporting said the backdoored version had been available for nearly 12 hours, and Cursor urged affected users to reset credentials while expanding warnings to two versions.

That did not stop product shipping. On 5 Nov Cursor introduced the ability to build live apps in Cursor, including automatic database and authentication setup. On 7 Nov it introduced voice chat for communicating with agents across the Agents Window, IDE, web and iOS. The juxtaposition is the point: richer agent interaction is arriving at the same time marketplaces, extensions and credential boundaries are under pressure.

Vercel leaned into the same trust vector from the platform side. On 4 Nov it introduced Vercel Security Review in public beta, automatically reviewing every push to GitHub repositories for vulnerabilities and allowing the Vercel Agent to open fixes. On 5 Nov and 6 Nov, reports covered its roll-out of device-bound session protection against stolen session cookies, and on 7 Nov TechRepublic reported Vercel's DBSC protection against session cookie theft.

GitHub also framed security as a core workflow issue. On 30 Oct it announced an npm Firewall to prevent installation of malicious packages, and on 6 Nov it upgraded secret scanning detectors. Lovable, meanwhile, was pulled into the same conversation on 6 Nov, when two articles reported that applications built on Lovable AI had exposed sensitive user data, including passport details and 200,000 chat messages.

Vercel's fortnight was unusually dense. On 8 Nov its AI Gateway added support for DeepSeek V4.1 Flash 1017 Ultrafast, callable with one API key, without a provider account, with spend tracking and request traces. On 7 Nov the AI Gateway added DeepSeek Search and Fetch tools, while Chat SDK agents gained the ability to stream generative UI with json-render to platforms such as Slack and Teams.

The surrounding platform pieces also moved. On 6 Nov Vercel Schedules entered public beta, and Token Pass credit began applying automatically to model usage through AI Gateway regardless of authentication method. On 5 Nov the Workflow SDK added idempotent starts to prevent duplicate runs, and Kimi K3.1 Pro expanded on the AI Gateway with more Zero Data Retention providers and automatic failover. On 31 Oct Data Cache began moving to Runtime Cache for faster performance under load, more cache storage per project and better observability.

The hiring and customer signals matched the product direction. On 4 Nov Vercel added an OpenEvidence case study describing medical search for 650,000 doctors. The same day it listed three San Francisco hybrid machine-learning engineering roles, including Agent Auth, Compute and Connect. On 5 Nov it added four startup-focused roles in New York City and San Francisco. This was not a single AI feature push. It was a coordinated move around inference access, scheduling, workflow durability, security review, cache behaviour and go-to-market.

The quieter signal was the choreography of websites, navigation and ads. On 31 Oct Cursor added 422 product pages, three documentation pages, one blog page and one pricing page. On 1 Nov Replit changed its navigation to emphasise Agent, Apps, Websites and Landing Pages, while removing the previous Apps & Websites, Agents, Engineers and Operations sections. Replit also removed 58 blog pages on 1 Nov, 21 more on 2 Nov and 11 on 31 Oct.

Vercel repeatedly rearranged its surface area. On 31 Oct it launched two Next.js template product pages and removed an engineering manager careers page and an OrderCloud marketplace page. On 1 Nov it added four more Next.js template pages and removed three documentation pages and one pricing page related to eve. On 3 Nov it added nine new product pages focused on templates and a connection product, while removing nine blog pages related to AI and database management guides.

Lovable did the same, but with more marketing pressure. On 30 Oct it added one blog page, three careers pages and 16 product pages, while removing four careers pages, 14 product pages, one docs page and one blog page. On 31 Oct it added listings for Product Partnership Manager and Trust Communications. On 7 Nov it had 14 new Google ad creatives, with at least 40 creatives running.

Individually, these edits are easy to dismiss. Together, inside one fortnight, they show a category narrowing its language around app shipping, agents, templates and trust. GitHub's own surface shifted on 4 Nov, adding call-to-action buttons for Sign up, Download GitHub Copilot app, Subscribe and English, while removing buttons for Products, Developers, Businesses and Resources. The front door is becoming a product strategy document.