Cursor Launches Its Own Code Hosting Platform Origin — the Code Collaboration Layer Starts Being Rebuilt for AI Agents
The biggest story of the day is in the code collaboration layer: Cursor officially launched Origin, its own code hosting platform that reimplements the core of GitHub and is exp…
The biggest story of the day is in the code collaboration layer: Cursor officially launched Origin, its own code hosting platform that reimplements the core of GitHub and is explicitly designed with AI agents as primary users. On the same day, GitHub suffered a large-scale outage, with error rates of roughly 20% on the web and API traffic and close to 50% on Archive and Raw downloads. Taken together, the two events pushed the community to seriously ask: when the people writing code increasingly become agents, should hosting platforms be redesigned? Meanwhile, Stripe was reported to be acquiring OpenRouter for more than $7 billion, NVIDIA announced a partnership with SB Energy to lock in power capacity in Ohio for dedicated AI factories (with OpenAI as a tenant), DeepSeek Harness saw rapid plugin ecosystem growth after going open source, and OpenAI — while unlocking a 1M-token context window for Codex — was reported by the Financial Times to have disbanded the team responsible for assessing “catastrophic risks.”
One: Cursor launches Origin code hosting, same day as the GitHub outage
Cursor today opened Origin, its own code hosting service, to all paid users in early beta. Based on the official changelog and community analysis, Origin is not a Cursor skin over GitHub: it reimplements repositories, commits, pull requests, code review, branch protection, permissions, and other core hosting features. Users can create repositories with the cursor.com/codebase/ prefix or mirror an existing GitHub repository into Origin, with pull requests synced bidirectionally; in mirror mode GitHub remains the source of truth, but users can “Detach from GitHub” at any time, making Origin the new source of truth. GitHub Issues, Actions workflows, and secrets are not synced.
Origin’s technical foundation comes from the Graphite team Cursor acquired in late 2025, which built stacked-PR management that naturally fits parallel agent development. In the Compile conference demo, Origin reached 22.6 commits per second in a single repository, 296,000 clones per hour, global sync latency below 400 ms, and built-in AI-driven automatic merge-conflict resolution. Cursor has said that 35% of its internal merged PRs are already created autonomously by agents running in cloud VMs — the numbers point to a scenario where a dozen agents open branches, commit, and file PRs in one repository simultaneously, which turns a hosting platform designed around sequential human-developer workflows into a bottleneck.
On the same day, GitHub confirmed a large-scale incident: web and API traffic error rates around 20%, Archive and Raw download error rates near 50%, and degraded performance across Actions, API, Pull Requests, Issues, and Webhooks. The community juxtaposed the two events with a mix of jokes and serious commentary: “when GitHub goes down, developers worldwide get an enforced day off,” and “in the agent-coding era, GitHub has become half of the AI infrastructure.” To be clear: there is no evidence of causation between the GitHub outage and Origin’s launch — they simply happened on the same day.
Evidence boundary: Origin’s feature list and performance figures come from Cursor’s official release and community summaries — company claims. GitHub’s outage data comes from official status updates as relayed. Both stories are consistent across multiple sources, but the performance numbers have not been independently verified by third parties.
Sources:
Two: Stripe reportedly acquiring OpenRouter for more than $7 billion
AI Valley reports that Stripe has agreed to acquire AI infrastructure company OpenRouter for more than $7 billion, just months after OpenRouter was valued at around $1.3 billion. OpenRouter is a unified API gateway aggregating 400+ models and serving roughly 8 million developers, letting users switch model providers through one interface — using whoever is cheaper or more reliable. For Stripe, the deal looks like a bet on owning the “infrastructure tollbooth” of the AI application layer: money spent on model calls, and the commerce around those calls, could flow through Stripe’s stack.
The direct impact on developers is unclear. OpenRouter’s value as a neutral aggregation layer rests on not favoring any single model vendor; after being acquired by a payments giant, its independence and pricing strategy are open questions. The news currently rests mainly on a single report (AI Valley); hubtoday’s summary also says the deal has been finalized, but no official announcement from either company has been seen. The valuation multiple (about 5.4x within a few months) itself reflects the market’s pricing of AI infrastructure access.
Evidence boundary: the acquisition price and valuation are media reports, not confirmed by Stripe or OpenRouter.
Sources:
Three: NVIDIA locks in Ohio power capacity with SB Energy, OpenAI as tenant
Jensen Huang announced on X that NVIDIA is partnering with SB Energy to secure power capacity (LPS) at the PORTS-Pike technology park in Ohio, dedicated to NVIDIA AI factories, with OpenAI as a tenant. NVIDIA’s official blog published the same news. According to Huang, the initial deployment is expected to provide 4.25 GW of AI factory capacity, roughly 1.5 million NVIDIA GPUs per generation, corresponding to $150–200 billion in revenue potential; OpenAI has committed to deploying about 12 GW of NVIDIA compute by 2030, expandable to 16 GW, with a total opportunity of roughly $600 billion.
A comparison figure was widely cited in community discussion: one researcher noted that this single Ohio site (on the order of 8 GW) is 8.6× the capacity of the EU’s entire seven-gigafactory “AI Continent” plan (0.93 GW). Whether or not the comparison is exact, it points to a clear trend: compute expansion is shifting from “buying GPUs” to “locking in power,” and power capacity itself is becoming a critical asset in the AI race.
Evidence boundary: all capacity, GPU count, and revenue figures come from NVIDIA/Jensen Huang — company claims; OpenAI’s tenancy commitment likewise originates from that announcement. The EU capacity comparison comes from a single researcher’s post.
Sources:
- https://x.com/JensenHuang/status/2089331487342829862
- https://blogs.nvidia.com/blog/securing-the-infrastructure-of-intelligence
Four: DeepSeek Harness ecosystem explodes after open-sourcing, API prices rise in parallel
DeepSeek open-sourced DeepSeek Harness (DSH), the tool its internal engineering teams use, earlier this week, and the community response has far exceeded typical open-source launches: GitHubDaily says it has “passed 150,000 stars within days of release,” plugin count exceeded 1,000 within days, and the community has assembled an awesome-dsh-plugin collection plus a companion plugin marketplace; the Feishu (Lark) plugin also shipped a visual-configuration 0.2.0 release. The “everything is a plugin” design (even memory is a plugin) lets developers quickly assemble their own agent products. Some practitioners believe “DSH-based agent products can now be built easily,” while others caution that much of the plugin ecosystem is still toy-level and needs a sustainable paid model long term.
In parallel with the open-sourcing, prices went up on the commercial side. Multiple community posts report significant DeepSeek API price increases, with some users citing up to 12× for certain tiers; hubtoday’s summary discloses a peak/off-peak pricing arrangement: peak-hour invocation fees have doubled, and running during off-peak hours can significantly reduce costs. After the increase, some users shifted to alternatives such as Xiaomi Mimo, while others complained “raising prices is fine, but don’t cut quality.” DeepSeek V4 Flash is currently free for a limited time, which some developers read as a market-grabbing move.
Evidence boundary: community figures such as 150k stars and 1,000+ plugins are not officially confirmed; different sources disagree on the magnitude of the price increase (2× peak vs. 12× claims) — the official confirmation only covers the existence of peak/off-peak pricing.
Sources:
- https://x.com/GitHub_Daily/status/2089502559816048647
- https://x.com/AYi_AInotes/status/2089285976178794586
Five: OpenAI’s push and pull — 1M-token Codex context, disbanded safety-evaluation team
On the Codex side, Tibo — the OpenAI lead for Codex — exposed a “full-power switch”: three lines in config.toml (model set to gpt-5.6-sol, model_context_window set to 1000000, model_auto_compact_token_limit set to 900000) unlock a 1M-token context window for GPT-5.6 Sol; community tests show the Codex desktop app actually reaches roughly 820K of usable context. Multiple posts warn that “available” does not mean “you should use it”: the default window is tuned for performance and cost, and most of a 1M-token context would be old plans, tool-call logs, and stale assumptions — “attention capacity is not attention quality.” Around the same time, GPT-5.6 Sol went 50% off on OpenRouter and Vercel’s AI Gateway (through September 18, 2026, both standard and fast tiers), widely read as inventory clearing ahead of a model-generation transition. Tibo also mentioned that Codex will get Astra (widely assumed to be GPT-6), and several practitioners predict Astra will natively support long-horizon multi-agent work.
On the governance side, the Financial Times reports that OpenAI shut down the team responsible for assessing whether its models could pose “catastrophic risks.” The team was formally disbanded at the end of July, with biological and cybersecurity evaluation duties moved into existing departments. HubToday’s summary confirms the change, calling it “governance turbulence widening ahead of the IPO.” Read together, the two stories frame a tension the community is discussing: on the product side OpenAI is pushing context and agent capabilities to their limits as fast as possible, while on the governance side it is contracting independent evaluation capacity.
Evidence boundary: the 1M-context configuration comes from Tibo’s own post and multiple user tests — credible first-hand information; Astra is rumor, not officially confirmed; the team disbandment comes from an FT report with no public OpenAI response.
Sources:
- https://x.com/aiwarts/status/2089353226474946852
- https://x.com/daniel_mac8/status/2089379592188436865
Six: WeCom opens office capabilities to AI agents; Tencent and ByteDance bet on the same direction
WeCom (Enterprise WeChat) launched an AI capability page that provides official capabilities to AI scenarios via CLI and MCP. Listed agents already include DeepSeek Harness, Codex, Kimi Work, MiniMax Code, WorkBuddy, and enterprise self-built agents; available capabilities cover messaging, email, documents, online spreadsheets, smart tables, smart docs, to-dos, calendar, meetings, WeDrive, and contacts. External agents can now legitimately operate real office data inside WeCom rather than relying on unofficial interfaces.
Community analysis links Tencent’s recent moves: on the consumer side, WeChat is pilot-testing its native AI agent “XiaoWei,” which can operate WeChat directly via text or voice (sending messages, making calls, changing settings, invoking mini-programs, even generating a mini-program from one sentence); on the business side, WeCom is opening its office capabilities to all external agents. ByteDance, meanwhile, merged the Feishu product team into the Doubao product team at the end of July, with Feishu GTM folded into the Volcano Engine organization. The shared judgment across both companies: the next stage of competition is no longer just about models — it’s about who can connect agents to the most real-world tools, data, contacts, documents, and workflows. The office software entry point is shifting from “a row of buttons” to “one sentence of instruction.”
Evidence boundary: details of WeCom’s open capabilities come from a single community long-form post (MaxForAI); the WeChat “XiaoWei” pilot and ByteDance reorganization are relayed in the same post, without cross-verification against official announcements.
Sources:
Seven: The Qwen3.8-27B local-deployment wave — community benchmarks diverge
Days after Qwen3.8-27B’s release, Hugging Face already shows 500+ quantized variants and 100+ fine-tunes. Multiple people report impressive local numbers: demos claim 200 tok/s on a single RTX 5090, or “over 100 tok/s,” leading some to conclude “a frontier model from six months ago now runs on a gaming GPU”; others say Qwen 3.8’s agentic capability beats DeepSeek Pro. One developer joked that “buy a GPU, deploy it, and you’ll have unlimited tokens.”
But a second set of tests offers counter-evidence. Jina AI’s Han Xiao found that qwen3.8-27b strongly prefers bash/grep in agent tasks: across 37 hard search queries on private data, semantic search calls dropped 33% and bash calls rose 133% — while accuracy actually improved (4/37 wrong vs. 3.6-35b-a3b’s 11/37). Another user noted that with the default reasoning effort set to xhigh, the model severely over-engineers: drawing a pelican riding a bicycle as an SVG took 21 minutes, and drawing a simple circle produced an animated, gradient, concentric-circle monstrosity — “xhigh is the opposite of fail fast.” Han Xiao himself also cautioned that many of today’s “local 5090 Qwen 3.8” speed numbers may be inflated by autoresearch processes, noting that “a casually designed autoresearch usually ends in reward hacking.” Separately, the community discovered that a previously listed “Qwen3.8-35B” model does not exist — it was a staff typo that has since been removed.
Evidence boundary: all performance numbers come from individual tests and demos, they contradict each other, and none should be treated as a benchmark conclusion; Alibaba has published no local-inference benchmark. The 27B model itself is an Alibaba release.
Sources:
Eight: Unitree lists on STAR Market tomorrow — “first humanoid-robot stock”
Unitree announced it will list on Shanghai’s STAR Market on August 19 (Beijing time) at an issue price of RMB 150.80 per share, implying a market cap of about RMB 60.99 billion, with expected proceeds of roughly RMB 6.1 billion. Prospectus figures: revenue of RMB 159 million, 393 million, and 1.699 billion for 2023, 2024, and 2025 respectively, with net profit of -RMB 11.15 million, 95.47 million, and 278 million — one of the few profitable general-purpose robotics companies. Around the IPO, community storytelling centers on sales chief Chen Li, who left Hikvision to join Unitree in 2018 when the company had only about RMB 100,000 in the bank, and on expectations that the “first humanoid-robot stock” could push toward a RMB 100 billion market cap.
The same day, specs for a new Unitree robot circulated in the community: 2-meter standing jump distance, top speed of 12.66 m/s (roughly 7.9 seconds for 100 meters), and 0.85-meter legs. These figures come from relayed demo reports, not an official spec sheet. The AI relevance here is more about embodied-intelligence targets entering the secondary market: the A-share robotics sector now has an anchor company that is actually profitable on product sales.
Evidence boundary: issue price, market cap, and financials come from IPO-related reporting with high credibility; robot specs are relayed; the “RMB 100 billion market cap” is community expectation, not fact.
Sources:
High-value briefs
- Cumora open-sourced — AI agents as full members of your chat group: yetone open-sourced Cumora, a Slack-like interface where the “colleagues” are all AI (Atlas the researcher, Bram the engineer, etc.) — with names, personas, memory; they can DM, form groups, claim tasks, and send and receive real email, and may proactively speak up without being @-mentioned. Two run modes: Cumora Cloud (agents in managed Kubernetes pods, driven by OpenAI’s Responses API) and BYOA (Bring Your Own Agent), where the agent’s “brain” is your local Claude Code or Codex CLI and keys never pass through Cumora’s servers. The multi-agent coordination design is worth attention: replies based on stale info get intercepted, task claiming is atomic, and a lightweight “little brain” triage layer decides whether to wake the big model. Community response: “multi-agent collision treated as a first-class problem.”
- Meituan reviews its company-wide “shrimp farming” AI experiment: per Phoenix New Media’s report of an internal talk, Meituan’s core local-commerce CEO Wang Puzhong reviewed the February–March company-wide AI transformation (internally called “shrimp farming”): the result was soaring bills — millions of RMB per day — plus AI-generated errors interfering with real operations. A rare high-level cost post-mortem of large-scale AI trial-and-error, standing in contrast to the “AI lowers dev costs” optimism.
- 404 Media tracks Amazon’s bulk book buying — scanned for AI training, then destroyed: 404 Media planted trackers in rare books and revealed an undisclosed Amazon operation: bulk-purchasing books, scanning them for AI training data, and destroying them; tracking shows the books ended up at an Amazon AI training facility. It mirrors the earlier Anthropic book controversy; the copyright/training-data boundary issue continues to heat up.
- The “Niulai” phenomenon and its meme coin: an animated film hand-built over five years by a mother-and-son team in Dalian went from viral mockery to box-office reversal (daily gross jumping from RMB 921 to 6.09 million, cumulative past 10 million), while its namesake meme coin hit a peak market cap of $46.79 million in three days before pulling back to ~$33.51 million. At least seven same-name contracts exist on-chain; the CZ “burn” was debunked (a fake contract forged the destruction); the liquidity pool is only ~$1.19 million with a market-cap-to-liquidity ratio over 28:1. A-share stocks like Luoniu Mountain hit limit-up and brokers issued risk warnings. It’s a complete sample of an “attention machine”: from trending topic to contract deployment in under 24 hours, with derivatives briefly worth 30× the underlying asset.
- llama.cpp author’s “Inception Mode” trick: ggerganov shared that when a model thinks too long without acting, you can plant a thought directly into its reasoning — “am I overthinking? Go gather more task info first” — and the model often stops spiraling and starts calling tools. Commentary: prompt engineering is shifting toward studying “what to inject mid-thought to make the model take a better next step.”
- Claude Code performance and deployment updates: the Claude Code CLI now uses 2× less CPU at p99 (fixing Bun’s garbage collector stealing CPU mid-turn); an official case study says ABC Legal deployed 50+ Claude Managed Agents that cut costs by up to 50% for certain legal tasks, with a feedback loop for continuous improvement. Cost figures are Anthropic’s company claims.
- Agent-skills research paper: skills stabilize execution rather than inject knowledge: a study across 8,135 normalized trial records finds that when skills help, 65.7% of cases involve “procedural anchoring” (stabilizing execution) and only 4.5% involve explicit knowledge injection; expanding the skill pool from 5 to 100 drops actual-use precision from 29.6% to 3.3%. A timely reminder for developers piling on skill packs.
- Hermes Desktop Bot Mode goes live: a community post says Hermes Desktop’s Bot Mode is now live — each bot can have its own model, memory, skills, tools, and persona, with bot-to-bot collaboration. Source is an X user post, not an official announcement.
- Google open-sources zero-trust AI agent example: an ADK + Gemini customer-service/returns agent demo that defends against prompt injection via three hard security layers outside the LLM context (hardware-backed signatures for non-repudiation, gVisor sandboxing for dynamic code, a deterministic semantic gateway validating business logic), stating plainly that “system prompts are soft constraints and cannot be the security boundary.”
- Remotion Skills 2.0: video-generation skills moved to a “router + 12 specialized sub-skills” architecture, added AI-generate-then-visually-edit-in-Studio with edits written back to source, and dropped built-in aesthetic preferences in favor of full user-prompt control. A concrete case of skill engineering.
- The AGI Bar makes Reuters: a Beijing Zhongguancun AI-themed bar named “AGI” (registered Chinese name “Knowledge Distillation,” a pun on both liquor distillation and model distillation) was covered by Reuters; two DGX Spark units inside give customers free unlimited DeepSeek tokens for coding via Wi-Fi, and its signature drink “AGI” sells for RMB 9.9. The bar reportedly runs inventory, reservations, utilities, and membership on AI agents, and opened a Shanghai branch in June.
- NVIDIA open-sources NOOA: NVIDIA Object Oriented Agents, described as a Pythonic framework for building AI agents, per community relay.
🕐 Selected hourly signals
| PT time | Signal | Why it matters |
|---|---|---|
| 06:20 | Jensen Huang announces NVIDIA–SB Energy Ohio deal with OpenAI as tenant | Power capacity becomes a critical AI asset; single post drew 1,300+ likes |
| 10:20 | OpenAI’s Tibo asks X what Codex is “obviously” missing | Codex lead publicly soliciting direction — 450 likes, 720 replies; hints at near-term product moves |
| 07:38 | Demo: Gemini 3.7 Flash plays Wordle on an Android emulator via ADB | Multimodal agent mobile-control scenarios; latency and visual reasoning are the differentiators |
| 06:03 | GPT-5.6 Sol at 50% off on OpenRouter and Vercel AI Gateway | Frontier-model discounts becoming routine; another inventory-clearing signal ahead of a generation shift |
| 07:00 | DeepSeek V4 Flash free for a limited time | Customer-acquisition move alongside the price increase |
| 15:00 | Community finds “Qwen3.8-35B” doesn’t exist — staff typo, removed | Model-info chaos reflects release velocity; community still wants a 35B |
| 05:40 | A 6-plugin DeepSeek Harness roundup (including a visual marketplace entry) | DSH ecosystem moving from “has plugins” to “has plugin management” |
| 13:20 | Claude Code 2.1.234 about to be released | Claude Code iteration continues at high frequency |
| 08:40 | Santiago quips “first they stole our data, sold it back, now they watermark it” | Text/image watermarking controversy heating up; one of the day’s most-shared threads |
Editorial conclusion
Today’s main thread is not a model release but accelerating reconstruction of the AI infrastructure layer: code hosting being rewritten for agents (Cursor Origin), a model-aggregation gateway bought by a payments giant (Stripe/OpenRouter), power becoming the new bottleneck in compute competition (Ohio), and domestic office software opening real data to agents (WeCom). In the same window, OpenAI unlocks 1M-token context while shrinking independent safety evaluation, and DeepSeek open-sources Harness while sharply raising prices — open-sourcing and commerce, aggression and contraction, coexisting inside the same companies. Two takeaways for readers: first, the agent-era competition is moving down the stack from models to tools, data, hosting, and energy; second, community benchmarks (especially local-inference speeds) are noisy — check the evidence before citing them.
Sources and method
This edition reviewed all 29 raw capture files in the 2026-08-17-pt directory (3 named sources + 20 hourly captures), deduplicated into roughly 25 candidate signals, of which 14+ carry substantive evidence — classified as a rich signal pool. Main limitations: most product and financial figures come from company statements or a single outlet and are flagged with evidence boundaries; local-model performance numbers contradict each other and were not cited as conclusions.
