AI coding moves from magic to margin control

The category is no longer selling autocomplete

The strongest signal this fortnight was the move from assistant features to operating systems for software work. On 12 Sep, Cursor introduced Projects, a feature for larger tasks such as features or full applications, with persistent context and delegation to multiple subagents. On 15 Sep, GitHub released copilot-sdk v1.0.14-preview.1 with pause and resume for durable factory runs at checkpoints. Vercel, meanwhile, shipped around agentic infrastructure: on 14 Sep it released eve 0.54.5 with changes for self-modifying agents and Slack channel routing, and on 19 Sep it added experimental WebMCP support to mcp-handler.

The common thread is clear. The market is trying to make AI coding less like a clever editor and more like managed production infrastructure. That makes the buyer more operational. It also raises the bar. Once tools handle longer-running work, budgets, permissions, sandboxes, storage, model routing and incident response become product features, not back-office concerns.

Lovable’s 19 Sep acquisition of Sutro, the company behind the SLang programming language, fits the same arc. The reported rationale was AI software transparency. In a market where builders can generate and ship quickly, explainability and inspectability are becoming part of the core product promise.

Control became the week’s most important feature

Vercel had the clearest control narrative. On 14 Sep, it increased Sandbox storage to 64 GB per sandbox and added permissions for connector management on Pro and Enterprise plans. On 17 Sep, it said Hobby plans would automatically delete inactive deployments for accounts exceeding the 10 GB limit, and that Secure Compute and Static IP builds would start 64% faster through prewarmed build containers. On 19 Sep, it added Enterprise spend management budgets with email alerts, webhooks and the option to pause production deployments.

GitHub moved along the same axis. On 14 Sep, its blog detailed new Copilot capabilities for configuring cost and quality in auto model selection. On 17 Sep, it announced general availability for Copilot budget increase requests, alongside security and access-management updates including automated SSO authorisation for classic PATs and SSH keys.

Pricing also showed the category shifting from acquisition to usage governance. On 13 Sep, Replit removed its Starter tier and reshaped Free Mode, Core and Pro benefits. On 18 Sep, it cut Core from $20 to $18 and Pro from $100 to $90. Cursor added a 20% annual billing promotion on 15 Sep. Vercel added included allowances and metered items on 14 Sep and 19 Sep. These are not cosmetic price-page tweaks. They are the mechanics of a market learning that AI coding demand is real, but compute, storage and model usage must be bounded.

Cursor had the loudest week, and the messiest one

Cursor dominated attention. Multiple reports on 20 Sep said Cursor had been acquired by SpaceX, with terms not disclosed in that event, while a 15 Sep funding item separately reported SpaceX had closed a $60 billion acquisition of Cursor. On 12 Sep, Cursor’s MCP Server was reported as powered by ZoomInfo as a go-to-market context layer. On 18 Sep, it listed five roles across engineering and sales, including Software Engineer and AI Deployment Manager, after four job openings had appeared on 13 Sep.

The product surface was also rewritten at speed. On 18 Sep, Cursor added six documentation pages and 94 product pages while removing two documentation pages and 98 product pages. On 20 Sep, it added another 16 product pages and removed 15 product pages plus one documentation page. That is the sort of churn that matters more than a launch post. It suggests repositioning under way, and it happened while the company was also appearing in press coverage around SpaceX, VS Code comparisons and enterprise use at Grab.

Reliability was the counterweight. On 15 Sep, Cursor reported a minor Grok service degradation lasting about 53 minutes. On 18 Sep, it reported five service-degradation incidents, including two major Grok Bot incidents lasting about 124 minutes and 390 minutes. On 20 Sep, it reported three more incidents, including two major Grok Bot degradations lasting about 343 minutes and 176 minutes. The market will forgive fast-moving infrastructure, but enterprise buyers will notice repeated agent and bot degradation during an ownership and positioning reset.

GitHub’s advantage now carries a visible trust tax

GitHub’s distribution machine kept moving. On 11 Sep, its customer page added logos from Pinterest, DAZN and Meta, plus case studies for Pinterest, SPH Media, DAZN, Meta and Front. On 15 Sep, it added Wealthsimple, Itaú, Plaid and The Hartford logos, with case studies for Wealthsimple and Itaú. On 17 Sep, Zendesk, WEX and Wayfair appeared, with Zendesk and WEX case studies. On 20 Sep, GitHub added a Tray.io case study for GitHub Team.

The product cadence was equally heavy. GitHub launched AI Scan for pull request APIs in public preview on 11 Sep, published Copilot updates on 14 Sep, released copilot-sdk v1.0.14-preview.1 on 15 Sep, and on 17 Sep described the migration of the GitHub Copilot runtime to Rust. On 18 Sep and 20 Sep, it created or updated Copilot-adjacent repositories including spec-kit-copilot, actions-migrations-via-copilot, copilot-language-server-release, gh-aw-firewall, codeql-coding-standards and vscode-codeql.

The risk column was just as visible. On 13 Sep, a GitHub Enterprise Server SSRF CVE with CVSS 7.4 HIGH was observed. On 15 Sep, CVE-2026-19118 described a GitHub Enterprise Server time-of-check time-of-use race condition with CVSS 7.5 and remote code execution by an authenticated user, alongside another September CVE. On 21 Sep, five CVEs were observed that specifically mention GitHub, including issues relating to command injection and insufficiently protected credentials in GitHub Copilot and Visual Studio Code. GitHub also reported a critical incident on 14 Sep lasting about 89 minutes, a major Gemini 3.8 Flash degradation on 17 Sep lasting about 628 minutes, and a Copilot OpenAI model error incident on 20 Sep lasting about 50 minutes. The point is not weakness. It is scale. The largest platform in the category is also the most exposed surface.

The growth play moved into templates and proof

Vercel’s homepage was not the story. Its sitemap was. On 11 Sep, Vercel added 41 product pages tied to templates and tools while removing 40 blog pages. On 14 Sep, it added 22 product pages and removed five documentation pages, nine blog pages and ten product pages. On 17 Sep, it added eight more product pages and removed AI Gateway model pages, a careers page and two blog pages. On 19 Sep, it added an AI Gateway virtual models documentation page and 16 product templates across Astro, Gatsby and Next.js.

Lovable showed the same pattern, but with a more explicit acquisition and localisation flavour. On 10 Sep, it added 30 product pages and new partner-program content while removing blogs, product pages, careers pages, a customer page and a docs page. On 17 Sep, it added blog pages on faster previews and a Salesforce partnership while removing nine careers pages and four project-management template pages. On 19 Sep, it added four blog pages about Sutro and 14 product pages for app and website builders in multiple languages.

Replit’s demand signals were more geographic and commercial. On 14 Sep, it opened its first international office in London and added partnership pages for Etisalat, Salesforce and Slack. On 15 Sep, its public customer wall added Payouts.com and Spell Strike Universe. On 16 Sep, it added a Spell Strike customer page and eight role-based product pages, while press reports cited $300 million in projected ARR for 2025 and a $9 billion valuation. The quiet lesson is that this market is no longer won only in the IDE. It is being won in search pages, customer walls, case studies, pricing tables and incident histories, often before the formal announcement lands.