Agent builders trade polish for control

The agent layer is becoming plumbing

The most important releases this fortnight were control-plane work, not coding theatre. On 8 Oct, Cursor introduced a way for users to see and interact with local agents running on their computers through the Cursor iOS app, with the feature available by default except on Enterprise plans. On 7 Oct, Vercel added confidence-based fallback model escalation to AI Gateway. On 10 Oct, Vercel said agents can buy domains directly through the Vercel CLI, while new Pro teams now default to 30 days of deployment retention.

GitHub kept moving in the same direction. On 6 Oct, it released gh-aw 0.91.2 with richer unified agent sessions and improved secret masking. On 8 Oct, gh-aw 0.91.5 added model routing and engine resilience work, with reliability fixes for Codex, Windows and safe outputs. On 10 Oct, gh-aw 0.91.7 introduced a native Agy engine, expanded the Git-backed work queue and improved model-routing visibility and security checks. On 12 Oct, gh-aw 0.91.8 added unified agent session updates, security enhancements and improved diagnostics.

This is the market’s centre of gravity now: not who can generate code, but who can route work, contain agents, inspect sessions and keep development systems auditable. GitHub’s 2026 patent publications, observed on 11 Oct, fit that pattern too, covering areas including scheduling continuous integration tasks, AI agent workflow instructions for web UI and vulnerability detection in open-source software code.

Replit’s repositioning became visible in the seams

Replit’s fortnight looked like a deliberate move away from its older education-heavy identity and towards business software creation. On 6 Oct and 7 Oct, multiple reports said Replit acquired Atta, a startup focused on AI analytics, data analysis and interactive charts. No deal value was provided. On 7 Oct, Replit removed 35 product pages related to educational offerings, plus three blog pages and one documentation page associated with Slack app support.

The next day, on 8 Oct, Replit added nine product pages focused on AI applications, including AI app builder, AI business software, AI game builder, AI software creation and AI workflow automation. That one-day gap matters. A casual read sees a website refresh. A closer read sees a product narrative being swapped in public: education pages out on 7 Oct, business AI use cases in on 8 Oct, with the Atta acquisition already in the news cycle.

Pricing language moved at the same time. On 7 Oct, Replit added "$20 towards more powerful models" to the Core tier and removed "$20 towards most powerful models" from the Core tier. It is a small wording change, but in this category small wording changes are often where gross-margin discipline first appears. Replit also listed cloud security roles on 7 Oct and mobile roles on 9 Oct, which puts platform hardening and mobile experience beside the new business software pitch.

Lovable bought reach, trust and enterprise shape

Lovable’s public motion was unusually broad. On 7 Oct, Lovable acquired Sutro to improve AI software transparency, and the same event set reported Lovable crossing $600 million in annual revenue. On 7 Oct, its careers page added Forward Deployed Engineer, Forward Deployed Engineering Lead and Performance Marketer for Paid Social roles. On 9 Oct, it added career pages for Data Infrastructure Engineer, Growth Engineer and Marketing Engineer, while removing pages tied to AI-scaled customer success and growth and marketing software engineering.

The advertising cadence was just as telling. Lovable was observed with 40 creatives running repeatedly across the period, with 29 new Google ad creatives noted on 5 Oct, 32 on 6 Oct, 28 on 7 Oct, 27 on 9 Oct, 28 on 10 Oct and 39 on 11 Oct. Several batches had earliest first-shown dates in late September. That is not a single campaign launch. It is sustained creative rotation while the company is also adding transparency, forward-deployed engineering and data infrastructure signals.

There was operational friction too. On 8 Oct, Lovable reported three status incidents, including a minor issue with creating projects from workspace chats, a minor issue with projects stuck in Recovering acceptance after a draft merge and a major issue where some users experienced errors. The public shape remains expansionary, but the operating burden is now visible: paid acquisition, enterprise delivery roles and production reliability are all moving at once.

Vercel kept rearranging the agent workbench

Vercel’s week was a study in rapid surface management. On 6 Oct, it released Next.js 16.4.0-canary.61 with caching improvements and build metadata recording, and created repositories named storage and sandbox. On 7 Oct, it released [email protected] and created repositories including vercel-py, terraform-provider-vercel, swr-site and nomad. On 9 Oct, it released [email protected] with a fix for streaming output timeouts and created chat, v0-sdk, sandbox and repository-dispatch. On 10 Oct and 11 Oct, further canary releases and new repositories followed.

The website changed almost daily. On 7 Oct, Vercel introduced seven product pages related to templates and configurations while removing seven others, including edge-config and growthbook pages. On 9 Oct, it introduced pages for llm-agent and a Shopify store template, while removing the agents product page and OAuth apps documentation. On 10 Oct, it introduced a Next.js blog starter kit and product pages related to Next.js and ratzilla, then on 11 Oct launched pages for moz and v0 while removing the same blog starter kit page and the ratzilla product page.

That churn is not noise. In a market where homepage taxonomy becomes sales enablement within days, the add-remove-add rhythm shows how quickly vendors are testing which concepts deserve a permanent shelf. Vercel is shipping infrastructure for agents, AI Gateway, sandboxing and templates, but the packaging is still being tuned in public.

Reliability is the tax on ambition

The same fortnight that showed agent infrastructure maturing also showed how fragile the layer still is. GitHub reported two incidents on 6 Oct: a major disruption affecting some services for about 105 minutes and a critical GitHub Actions incident for about 218 minutes. On 8 Oct, GitHub reported three critical incidents affecting services including Git Operations, Issues, Actions and Pull Requests, with the longest lasting about 223 minutes.

Cursor had repeated service degradation. On 6 Oct, it reported four incidents affecting Cloud Agents, Grok Bot and Security Reviewer Agents, including one major Cloud Agents and Grok Bot incident lasting about 159 minutes. On 8 Oct, Cursor reported an 86-minute minor Grok Bot degradation. On 9 Oct, it reported Opus 5.5 elevated errors for about 58 minutes and 429 errors for about 97 minutes. On 11 Oct, it reported two further minor incidents affecting Cloud Agents and Grok Bot.

Security pressure was also visible. On 5 Oct, reports said Vercel confirmed a KVM zero-day vulnerability related to VM escape and announced a $50,000 bounty for the researcher who reported it. On 6 Oct, Vercel was again covered in relation to a full VM escape vulnerability in Linux KVM. In the GitHub ecosystem, CVE-2026-75101 and CVE-2026-77912 were noted on 8 Oct for GitHub Enterprise Server, covering an authorisation bypass and stored cross-site scripting issue. On 12 Oct, CVE-2026-100746 was noted for Coolify versions up to 4.1.0, affecting a GitHub App Set redirect function.