Agents take the operating layer

The most important movement this fortnight was not another coding assistant prompt. It was the migration of agents into the operating layer. On 29 Oct, Replit announced Replit for Business for managing AI agent deployments and highlighted per-agent identity tracking for enterprise AI governance. On the same day, Vercel put Vercel CI into beta, allowing users to run all continuous integration tasks, not just builds, on Vercel, and introduced Box Mount so agents can read and write Box files through the filesystem in Vercel Sandbox.

Vercel kept pushing that line. On 27 Oct it added live process input for stdio MCP tools in Sandbox. On 31 Oct it said Data Cache is moving to Runtime Cache and added service bindings for Vercel Daemons. On 1 Nov, Vercel Labs open-sourced agent-desktop, a tool for AI agents to operate using the full capabilities of a computer.

GitHub supplied the clearest framing. On 2 Nov, COO Kyle Daigle said AI agents could become bigger consumers of GitHub services than traditional developers. That statement landed after GitHub launched Copilot Fusion on 30 Oct and reported, on 1 Nov, a sixfold increase in monthly commits to 7.4bn, requiring an additional 4m CPU cores and 160 petabytes of storage. The category is no longer only selling developer productivity. It is preparing for non-human usage at infrastructure scale.

GitHub’s week was unusually coordinated. On 29 Oct, its homepage headline changed from 'The future of building happens together' to 'Software still starts with you', while new calls to action for Products, Developers and Businesses replaced older Platform, Solutions and Open Source routes. The same day, GitHub’s pricing page added GitHub Copilot access and Agent governance to Enterprise, moved everything from Free into Team, and removed Repository rules and the Audit Log API from Enterprise.

That packaging matched the product news. On 30 Oct, GitHub announced an npm Firewall to prevent installation of malicious packages and released Copilot Fusion with three pricing tiers and a 30 percent promotional discount. On 1 Nov, GitHub said Copilot Fusion offers model selection and orchestration, with claims of lower API cost than Claude Code. The public message was simple: agents belong inside GitHub’s enterprise control plane, not beside it.

The customer page reinforced that enterprise posture without adding new logos. GitHub added or featured case studies for ADEO on 27 Oct, Via on 28 Oct, LambdaTest, Philips, Buffer, Cesium and Knock on 29 Oct, Aevi, KPMG, x15ventures, GM and AstraZeneca on 31 Oct, then Zalando, Via and Tray.io on 1 Nov. The emphasis was existing proof, not logo theatre.

Cursor had the sharpest reminder that extension ecosystems are now part of the attack surface. On 31 Oct, Cursor 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. On 2 Nov, further reporting said the backdoored version had been available for nearly 12 hours, with up to 5,000 users impacted, and Cursor urged users to reset credentials while expanding warnings to two malicious versions.

Service degradation was also a recurring theme. Cursor reported a major incident on 29 Oct in which multiple services were degraded for around 118 minutes, then on 30 Oct had incidents under investigation for Cloud Agent Degradation and VS Code Extensions degradation, alongside resolved Auto Review and Origin and Cloud Agents performance issues. On 1 Nov, Cursor reported a minor Billing and Usage Dashboard degradation and a resolved incident involving multiple degraded services for about 44 minutes.

GitHub’s status page was busy as well. On 29 Oct, a critical Copilot services disruption lasted around 75 minutes. On 30 Oct, a minor Copilot Cloud disruption lasted around 150 minutes. On 31 Oct, GitHub reported a Dependabot incident of about 150 minutes, a major incident affecting Actions, Pages and Webhooks for about 93 minutes, and a critical disruption impacting GitHub services for about 32 minutes. Replit reported platform and agent incidents on 31 Oct, a 49-minute agent image processing incident on 1 Nov, and a major incident impacting agents and workspaces on 2 Nov. Vercel also had to respond to security pressure, implementing emergency firewall rules for a Brotli decoder flaw on 27 Oct and for a critical Node.js remote code execution flaw on 29 Oct.

The launch calendar only told half the story. The quieter signal was how many vendors rewired their public surfaces at the same time. On 27 Oct, Vercel added 22 product pages focused on templates and removed 22 marketplace product pages. On 28 Oct and 30 Oct it again added knowledge base pages while removing Next.js template pages, then on 31 Oct launched more Next.js template pages and removed an engineering manager careers page and an OrderCloud marketplace page. On 1 Nov, it added four Next.js template pages and removed documentation and pricing pages related to 'eve'.

Replit’s public surface moved just as fast. On 28 Oct it added 25 documentation pages around external codebases, GitHub and GitLab integrations, and agent functionality. On 1 Nov its navigation shifted from Apps & Websites, Agents, Engineers and Operations to Agent, Apps, Websites and Landing Pages. Across 31 Oct, 1 Nov and 2 Nov it removed batches of blog pages, including posts on AI adoption, production databases, API documentation, AWS guides, deployment analytics, outbound data transfer limits and GitHub credentials exposure.

Cursor’s sitemap changes were almost a metronome. On 28 Oct it added 25 product pages and three documentation pages while removing 27 product pages and one documentation page. On 31 Oct it added 422 product pages, three documentation pages, one blog page and one pricing page. On 1 Nov it removed 426 forum pages, including 422 product-related pages, three documentation pages and one pricing page. On 2 Nov it added 422 product pages and four documentation pages covering task management, interface changes, GLM 5.3 and AWS Bedrock integration.

Lovable showed the demand side of the same cycle. On 29 Oct it had 32 new Google ad creatives, with the earliest first shown on 24 Oct, and 40 creatives running. On 1 Nov, another seven new creatives were observed, with 40 still running. Its careers and product pages also churned through roles in customer success, payments, forward-deployed engineering, data infrastructure, trust communications and partnerships between 28 Oct and 2 Nov. The point is not any single page. It is the simultaneity: product packaging, hiring posture and demand generation changed before most of the week’s formal narratives settled.

The pricing moves lined up with the product moves. On 29 Oct, Replit added even more Free Mode usage and $100 credits to Pro, while removing 10 parallel agents, more Free Mode usage to chat and create, and the ability to invite up to 50 viewers from Pro. That is a material reshaping of what the paid plan foregrounds: credits over parallelism and sharing limits.

GitHub’s 29 Oct Enterprise change made Copilot access and Agent governance part of the upper tier. On 30 Oct, Vercel added BotID at $1 per 1,000 Deep Analysis checks and included 360 GB-hours of Fluid Provisioned Memory per month in Pro. Lovable added a monthly credit balance to Pro on 30 Oct.

The pattern is narrow but important. The vendors in this set are not only repricing seats. They are exposing consumption units, security checks, memory allocation, credits and governance as commercial primitives. That fits the same week’s operational story: once agents touch CI, filesystems, deployments and package installation, the bill has to follow usage and control, not just human users.