If you tried to push, merge, or run CI yesterday afternoon and couldn't, that wasn't you — GitHub was down worldwide. And in the same 24 hours, the ground under your source control shifted twice more: Cursor shipped a GitHub alternative built for AI agents, and an AI-video startup posted a $400M raise on revenue that grew 35x in a year. Here's the whole edition in one screen:

The through-line: the platform layer you build on is fragile, contested, and richly funded all at once. None of it breaks your company by itself — but each is a reminder that every platform under you is temporary. Keep a second path wired. Here's what to check before lunch.

1. GitHub went down worldwide — and the damage reached far past github.com#

If your afternoon stalled on Aug 17, here's why. GitHub's outage opened at 13:40 UTC and ran for hours. GitHub reported that API Requests, Actions, Git Operations, Issues, Pages, Pull Requests, and Webhooks were mitigated at 16:59 UTC — about 3 hours and 19 minutes — while Copilot joined the affected list around 14:31 UTC and lingered as a Major Outage after the rest recovered, with some outlets clocking full resolution closer to eight hours because of Copilot authentication problems that outlasted everything else.

The number that matters for a founder isn't the duration — it's the error mix. During the incident roughly 20% of general web and API requests errored, and about 50% of archive downloads and raw-content requests failed. That second figure is why the outage broke things nowhere near github.com: npm install, pip install, Docker builds that pull raw files, and Go module downloads all lean on those raw and archive endpoints. A pipeline that looked independent of GitHub's uptime quietly wasn't.

What it means: Treat this as a free penetration test of your deploy chain. Most solo and small-team pipelines have a single point of failure at GitHub, and you find out during the outage, not before. Three cheap hedges pay for themselves: mirror your critical repositories somewhere else so your history isn't hostage to one host; cache your dependencies and build artifacts (a local registry proxy or a committed lockfile plus vendored deps) so an install survives a raw-content failure; and keep a manual deploy path you can run cold — a script that ships from your laptop when Actions is dark. If you already run agents in parallel worktrees, our guide to git worktrees for parallel AI agents doubles as a local-first resilience pattern, and checkpointing long-running agents to object storage keeps agent state off any single provider's uptime.

2. Cursor launched Origin — a code host built for AI agents, not humans#

The strategic counterpoint landed the same day. On Aug 17, Cursor launched Origin, a code-hosting platform built directly into its AI editor. It ships the table stakes — repositories, pull requests, code browsing, PR review and merge, two-way GitHub sync — plus CI and deploy integrations including Vercel, Buildkite, and Depot. What makes it notable isn't the feature list; it's the framing. Origin is pitched for "agent scale": workflows where AI agents open, review, and merge pull requests, not just people. It's in beta on paid plans, and it arrives weeks after Cursor-maker Anysphere became a SpaceX unit.

The launch-day overlap with GitHub's outage was, per Cursor, coincidental — and it plainly was; a code host with review, merge, and CI hooks is built over months, not shipped as an outage response. But the outage made it land harder, because thousands of developers spent the afternoon unable to push or merge and were, for a few hours, unusually open to alternatives.

What it means: The git-hosting layer is being re-contested for the first time in years, and the wedge is agents. If most of your PRs will soon be opened by an agent rather than a human, a host designed around that loop is worth understanding early — so spin up a throwaway repo on Origin and watch whether an agent-native flow actually changes your cycle time. But don't move your primary repos on day one of a beta: the entire value of git is that your history is portable, and the GitHub-sync feature exists precisely so you can evaluate Origin without burning that portability. For where Cursor sits against the field first, our Cursor vs Windsurf vs GitHub Copilot vs Claude Code comparison and the August coding-agent ranking are the two reads worth pairing with this.

3. Higgsfield raised $400M at $5.4B — on revenue that grew 35x in a year#

The capital side of the same day: AI video-and-image startup Higgsfield announced a $400M Series B at a $5.4B valuation, with DST Global leading and Goldman Sachs Alternatives' growth-equity arm, Intel Capital, Tribe, Smash, Fifth Wall, Valor, and others joining. The valuation is roughly 4x its ~$1.3B mark from earlier in 2026. The line that should stop a founder is the revenue: Higgsfield says annualized revenue reached ~$700M this month, up from about $20M a year earlier — a roughly 35x run-up — and that it now powers visual production for 390 of the Fortune 500, with more than 20 million content generations a month.

Unlike the day's other headlines, this one is an official company release, not a leak — the figures are the company's own, but they're on the record.

What it means: You can't buy the stock, so read it as a market instrument, not a trade. Two signals matter for anyone building in or near applied AI. First, an AI-native product with real distribution can compound revenue faster than almost any prior software category — $20M to $700M in a year is not a SaaS growth curve, it's a new one, and it resets what "fast" means when you're pitching. Second, late-stage capital is still writing nine-figure checks into applied AI at high multiples in August 2026, which is good news if you're raising and a warning if you're competing: assume a well-funded incumbent in any hot creative-tooling niche and win on focus and distribution, not feature count. For the fuller picture of where this year's AI money is actually landing, our August 2026 agent-funding roundup tracks the lanes the checks are flowing into.

Also on the wire#

A reminder on a story that's still landing: Anthropic began watermarking all Claude text, which we covered on Aug 13. Using Google DeepMind's SynthID-Text method, Claude models released after Aug 2 now embed invisible watermarks in generated text to meet the EU AI Act's transparency rules, rolled out globally with no regional toggle and, per Anthropic, no visible impact on quality, speed, or price; a detection API is "coming soon." If your product passes Claude-generated text off as human-written, that text is now detectable as machine-generated — factor it into both your positioning and your EU compliance. And the reported Stripe–OpenRouter acquisition we covered yesterday remains reported, not officially confirmed, with an announcement said to be expected this week — keep your model-router fallback wired until the terms are real.

The one move that covers all three: assume every platform under you is temporary. Mirror your repos, keep your git history portable, and compete on distribution rather than a feature moat — and an outage, a new code host, and a freshly funded rival become things you read about with coffee instead of things that set your afternoon on fire.