Coding tools move up the stack

The week’s defining event was Cursor’s reported acquisition by SpaceX. Multiple reports on 26 Aug, 27 Aug, 28 Aug, 29 Aug and 30 Aug put the deal at $60 billion, with Elon Musk publicly acknowledging the transaction and addressing Cursor engineers. On 30 Aug, reports said OpenAI would terminate its partnership with Cursor and cut off access to its models in November 2026.

Cursor did not respond by looking smaller. On 30 Aug it announced Origin, described as an agent-native alternative to GitHub. That followed a 27 Aug changelog update allowing users to begin projects from scratch without connecting to external source control, prompt the cloud agent immediately, and save work to a Cursor Origin repo. The direction is plain: Cursor is trying to own the code creation environment, not just sit inside one.

The supporting product edits were too closely timed to ignore. On 26 Aug Cursor added a Models navigation item and removed Review. On 27 Aug it added Grok Bot to navigation and removed Marketplace. On 27 Aug it released SDK bridge v1.0.30 with standalone executables and a stable sdk.v1 proto contract on npm and PyPI. On 27 Aug Grok Bot access was added to Pro, then on 28 Aug higher Grok Bot usage limits were added to Pro+ and the highest limits to Ultra. The company was re-pointing the front door, the pricing page and the developer surface in the same 48-hour window.

The risk is equally visible. Cursor had a Grok models degradation incident for about 54 minutes on 26 Aug, a Grok Bot degradation for about 42 minutes on 27 Aug, and a critical degradation affecting cursor.com and the IDE for about seven minutes on 28 Aug. It also faced press coverage on 27 Aug and 28 Aug about alleged abuse of Cursor Agent AI by threat actors. A company moving from IDE companion to agent-native platform now has to defend itself like infrastructure.

Vercel’s week was less theatrical and more commercially important. On 27 Aug it added Meta’s Muse Image and Google’s Gemini 3.5 Transcribe to the AI Gateway, both callable with a single API key, automatic fallbacks and spend tracking. On 28 Aug it added Claude Managed Agents with Chat SDK and brought Cursor into the AI SDK harness layer. On 29 Aug it added Tencent’s Hy4 Preview through the AI Gateway and enabled users to build and deploy eve agents directly from the dashboard. On 31 Aug it announced a 50 percent discount on MiniMax H3 and H3 Max video generation through the AI Gateway, applied to existing model IDs so current requests automatically receive the reduced pricing.

That is a coherent sequence. Vercel is not merely listing model integrations. It is bundling access, fallback, spend control, agent sessions and deployment into the same operational plane. Its 31 Aug chat package release, version 4.39.0, added Slack Agent Sessions lifecycle support, native stop cancellation through thread.signal, automatic session titles and compatibility with the legacy assistant_view path. The detail matters because agent products break at the seams, not in the launch post.

The content layer moved at the same time. On 29 Aug Vercel introduced six new product pages related to Next.js templates focused on AI and chat, while removing five pages related to AI models and one documentation page. That is a quiet but revealing swap: fewer standalone model pages, more use-case templates. The same day, it open-sourced vgpu for AI agent shaders and released Turborepo v2.10.13-canary.1 with a performance agent toolbox expansion.

Vercel also signalled enterprise operating intent. On 27 Aug it listed roles including SOX Manager and Strategic Sourcing Manager, and on 29 Aug roles including Sales Development Representative and Staff Data Platform Engineer. On 31 Aug CTO Malte Ubl announced a $1 million initiative to counter AI attackers. In a week when AI coding tools were linked to security incidents and model dependency shocks, Vercel’s public posture was gateway, governance and reliability.

GitHub’s answer was distribution and governance. On 27 Aug its blog announced general availability of a global model policy and said GitHub Apps can access enterprise billing data. On 30 Aug its blog detailed free secret scanning, while separate coverage highlighted Microsoft’s move of WinUI development to GitHub. On 29 Aug GitHub also announced the deprecation of GitHub Classroom and updates to upcoming Copilot policies and billing.

The customer page was busy. On 27 Aug it added logos including 3M, Caregility, Bolt, Ford and Procter & Gamble. On 28 Aug it added HOVER and Opn, with case studies including HOVER’s use of GitHub’s integrated development platform and Opn’s workflow automation and payment security. On 29 Aug it featured Mercari, Decathlon and ASOS, with case studies referencing GitHub Advanced Security, Actions, open source and Copilot.

The developer surface expanded in parallel. On 27 Aug GitHub’s top repository stars were reported as growing from 35,729 to 61,846 during the week, with new repositories including codeql, awesome-copilot, gh-stack, actions-lockfile and copilot-language-server-release. On 28 Aug it released gh-aw v0.87.8 with safe-outputs, MCP gateway reliability, security and dependency fixes, and created repositories including copilot-sdk and CopilotForXcode. On 29 Aug gh-aw v0.87.9 added expanded Windows runner support for Bash and PowerShell and enhancements for agentic workflows.

The blemish was availability. On 27 Aug GitHub reported a critical Actions incident lasting about 170 minutes, a minor Actions and Pull Requests incident lasting about 90 minutes, a Billing disruption under investigation and another minor disruption. On 28 Aug Copilot AI Model Providers had a critical incident lasting about 128 minutes. The strategic read is not that GitHub is weak. It is that once coding assistance becomes a model-mediated workflow, even the incumbent’s reliability envelope is tested in public.

Replit’s week was about audience broadening and model choice. On 26 Aug it cut Core from $25 to $20, replaced $25 towards the most powerful models with $20, and added a promo of up to 15 percent off. The same day it created a Bubble alternatives product page. On 27 Aug it launched a blog page on intelligent model routing, added a Growth Kit product page, and changed a website heading from Quality, Fast, Efficient to The best model for every task.

The next edits widened the buyer map. On 29 Aug Replit’s navigation added categories including Apps & Websites, Slides, Agents, Databases, Security, Founders, Product Managers, Engineers and Operations, while removing items including Agent, Database, Publish and Mobile. On 28 Aug it removed two product pages tied to dark mode use cases for operations and product managers. On 28 Aug it also listed Support Systems Lead roles in Foster City and New York, with descriptions focused on building systems to support scaling operations alongside product growth.

Lovable’s week was capital-intensive and content-heavy. On 26 Aug it completed a $400 million funding round intended for cost of goods rather than growth initiatives. On 27 Aug and 30 Aug further reports put the round at a $13.3 billion valuation. That framing matters in this category because cost of goods is the pressure point when users ask agents to generate, revise and deploy software at scale.

Lovable’s site activity pointed in several directions at once. On 27 Aug it added a Payroll Lead careers page, created 16 blog pages focused on design tools and processes, introduced a new product page, and removed several careers and blog pages. On 28 Aug it added an AI Ops Engineer careers page, published nine blog posts including enterprise features and integrations, and removed ten product pages related to educational and startup partnerships. On 29 Aug it launched multilingual presentation template pages and removed a corporate development lead careers page and four blog pages. The company also had minor incidents on 26 Aug, 27 Aug and 28 Aug, including degraded Lovable Cloud performance for about 134 minutes.

The fortnight was not quiet, but the most useful signal was not any single headline. It was the simultaneity of small edits. Cursor changed navigation, pricing, SDK packaging and source-control assumptions around the same moment it was reported to be losing OpenAI access. Vercel shifted from model pages to AI and chat templates while adding gateway providers, agent sessions and a video-generation discount. Replit cut price and rewrote its model-routing message while widening navigation from builders to functional buyers.

These are the moves operators usually miss because none is decisive alone. A removed Marketplace link, a new Models tab, a deleted AI model page, a rewritten heading and a live discount on existing model IDs are all minor in isolation. Together, they show the category moving from AI coding as feature to AI software production as operating system.

The competitive line is becoming clearer. Cursor wants the agent-native repository and IDE. Vercel wants the deployment, gateway and workflow layer. GitHub wants the trusted enterprise control plane. Replit and Lovable are pushing towards broader creation markets, where the buyer may be a founder, product manager or operations team rather than a full-time engineer. The market is still called AI coding, but the products are already fighting over who owns the application lifecycle.