Skip to content

Highlights

What happened at the companies we cover, by day, newest first. Every item links to its sources.

Showing 1–20 of 42 highlights

Thursday, Nov 12 42 stories

  • Modal becomes a founding member of LMSYS's new SGLang Consortium

    News
    NewsX

    LMSYS Org launched an SGLang Consortium on Nov 12 at the first SGLang Summit in San Francisco, and Modal is one of eight founding members alongside NVIDIA, AMD, Intel, Google, Meta, Cloudflare and RadixArk. Each member names a technical representative and joins two roadmap meetings a year, while SGLang's maintainers keep control of technical decisions and code acceptance. LMSYS lists Modal among the production users of the inference engine. Modal said on X that SGLang underpins some of its heaviest inference workloads and that its inference engineers were at the summit; the posts drew solid engagement and no pushback.

    Sources

  • Ambience launches Prepare, an agent that builds evidence-linked pre-visit chart briefings

    Launch
    News

    Ambience Healthcare introduced Ambience Prepare on Nov 12, pairing Chart Prep, which turns a patient's longitudinal record into an evidence-linked briefing before a visit, with Chart Chat for natural-language questions about the chart. Chart Prep runs on Ambience's Chorus context system, which lets the agent navigate notes, medications, labs and reports, reconcile conflicting sources and flag unresolved questions for the clinician. On a 46-case de-identified benchmark with three runs per case, Ambience reported its newer context harness surfaced 92.2% of clinician-ranked critical facts versus 78.7% for the production harness, with the model held constant. Retrieval-related misses fell from 19.9% to 9.3% in that comparison.

    Sources

  • Replit Agent can now buy supplies via Stripe Link and Visa and run Google Ads

    Launch
    XBlueskyNewsOther

    Replit gave its Agent two new jobs on Nov 12: buying business supplies through Link, Stripe's wallet, with Visa payment infrastructure, and creating and managing Google Ads campaigns. For purchases, a Browserbase-powered browser subagent fills the merchant cart and stops at checkout until the owner approves the total, with shell access and general browsing disabled during payment and approval checks enforced in code. Agentic Commerce is for Core and Pro users in the US and limited to one-time purchases. The ads feature drafts copy and images, recommends budgets and adds conversion tracking; Link promoted the launch on X, while some users asked how much spend they would trust an agent with.

    Sources

  • OpenAI temporarily extends Cursor's model access past the Nov 12 shutoff date

    News
    NewsX

    OpenAI updated its August notice on Nov 12 to say it has temporarily extended Cursor's access to its models beyond the proposed Nov 12 shutoff, with further updates to come. OpenAI originally said it would wind down its contract after SpaceX acquired Cursor's maker, citing concerns about SpaceX honoring its terms, and said it would not provide future models such as Astra to Cursor. A Taiwanese tech outlet checking the deadline found Cursor's model and pricing docs still listed GPT-5.6 models for paid tiers using prepaid credits, and advised users to test access in their own accounts rather than assume models were pulled. On X, developers flagged the deadline and asked how fast teams would switch models, but no final termination date has been announced.

    Sources

  • Devin Teams drops Desktop seat fees and adds free SWE-2 sessions, SSO and cost tiers

    Launch
    XNews

    Cognition upgraded the Devin Teams plan on Nov 12 so every team member can use Devin Desktop without buying extra seats. A new Free mode gives each user two concurrent sessions on Cognition's own SWE-2 model, and free SWE-2 in Desktop and CLI becomes a permanent part of Teams, Pro and Max. Admins get SSO via SAML or OIDC, per-user usage limits in up to three tiers with request-and-approve flows, and a standard BAA without an enterprise contract. Quota seats on Teams retire in 30 days and a startup program offers up to $65k in credits; reaction on X was positive, with several posts relaying the changes.

    Sources

  • Decagon launches Missions so its agents can pursue one customer goal for weeks

    Launch
    NewsXOther

    Decagon introduced Missions, which let its customer-service agent own a business outcome such as a prescription refill, small loan approval or lead qualification across many conversations, channels and systems over days or months. Teams state the goal in plain language, list the people to engage, and the agent picks each next step, including back-office actions like filling forms through Browser Actions while waiting on third parties. Each Mission keeps a running case record plus per-customer memory, and the logic is written as Agent Operating Procedures that teams can test and tune with Duet Autopilot. Decagon staff promoted the launch on X, citing mortgage lenders as an early use case.

    Sources

  • Abridge Assistant now surfaces each health system's own care pathways

    Launch
    NewsX

    Abridge announced on Nov 12 that its chart-aware Assistant can now pull a health system's institution-specific care pathways into answers alongside published medical literature. When a matching pathway applies, the Assistant shows the relevant section with a link back to the original document, leaving clinicians to judge how it fits the patient. Abridge says the Assistant is used across more than 300 enterprise health systems and that customers can turn the feature on through existing deployments that already passed security review. On X, clinicians and informatics leaders were mostly supportive, though several replies flagged that pathways drift out of date and asked who keeps versions current.

    Sources

  • Clay engineering stops line-by-line code review, shifts reviews to architecture

    Other
    Other

    Clay's engineering team published a Nov 12 post explaining that, because coding and review agents now catch most bugs, human reviewers no longer need to hunt for bugs or read every line on most pull requests. Reviews now concentrate on data schemas, infrastructure and library choices, service boundaries, new code patterns, and security and reliability risks. Clay wants design discussions to happen before a large PR is reviewed, and reviewers can ask for a short design write-up before looking at a complex diff. The company also plans a PR risk-scoring bot and is experimenting with auto-approval for low-risk changes.

    Sources

  • Report: Baseten looks to buy its own GPUs as rented inference capacity tightens

    News
    NewsXBluesky

    The Information reported that inference providers that resell leased compute are starting to buy hardware, and that Baseten is reportedly looking to buy GPUs; Fal and Together are named as further along. Per Dealroom's summary of the report, an unnamed source claimed CoreWeave stopped selling compute to Baseten after launching its own inference service, which someone close to Baseten disputes, and CoreWeave says its relationships with inference providers are good. Several points rest on unnamed sources, so this should be read as reporting rather than a confirmed plan. On X, the story was tied to the Fireworks breach as a sign that reliance on small neoclouds carries security risk.

    Sources

  • Vercel opens UI SDK beta: composable UI behaviors built for coding agents

    Launch
    XBlueskyOther

    Vercel released UI SDK in beta, an Apache-2.0 library that breaks interface components into reusable behaviors such as popups, selection, typeahead, roving focus, drag-to-reorder and virtualization. The core package ships with adapters for React, Vue, Svelte, Solid and Web Components, plus an installable agent skill so coding agents can assemble new components from 50+ behaviors. Vercel also claims agents can operate the resulting controls by reading the page instead of screenshots, at roughly double the speed and under half the cost. Developers on X and Bluesky shared the release widely, and reactions were mostly positive.

    Sources

  • Lovable adds scheduled Automations that run recurring agent tasks

    Launch
    Other

    Lovable's Nov 11 changelog introduced Automations, letting users ask in Chats for a task to run on a schedule, such as a daily check of published sites or a weekly signup summary. Lovable then posts its findings back into the chat. The feature is on Free, Pro and Business plans and each run is billed as chat usage. The same update removed auto-accept for workspace and project invitations, so each invite must now be accepted manually. Public reaction was limited so far.

    Sources

  • ACE Music v1 edges past Suno v6 on a blind vocal-music leaderboard

    Discussion
    NewsX

    ACE Studio's ACE Music v1 took first place on Artificial Analysis' blind-listening vocal music leaderboard with 1,160 Elo against Suno v6's 1,144, though Decrypt notes the confidence intervals overlap and Suno still leads the instrumental board. Decrypt's hands-on test found ACE gave more natural vocals and cleaner endings in mainstream styles, while Suno did better on niche genres such as melodic death metal. ACE's flagship is closed with no API, but the company plans to open-source a Flash version later in 2026. On X, Chinese- and Japanese-language posts spread the news widely, and some Suno users said v6 felt stale or produced repetitive intros.

    Sources

  • Bret Taylor makes the case for Poppy as the rulebook for personal agents

    Discussion
    NewsX

    Sierra CEO Bret Taylor gave an interview to retail newsletter The Aisle on Poppy, the Personal Agent Protocol that Sierra and Meta are developing with partners including Shopify, Stripe and Walmart. The protocol lets personal AI agents identify themselves to a business and lets the business set granular permissions, positioned as complementary to commerce protocols such as UCP. The discussion framed it against Amazon blocking outside shopping agents while Walmart opens a gateway. On X, at least one developer showed an early build bridging the Poppy spec to an existing auth provider, suggesting early hands-on interest.

    Sources

  • Harvey's Contract Intelligence now keeps full negotiation history and deal status

    Launch
    Other

    Harvey updated Contract Intelligence so each new counterparty markup is matched to the existing contract and reviewed alongside earlier rounds and related agreements. The AI review accounts for positions Legal already accepted, so any lawyer picking up a deal can jump to the open points. New contract statuses show whether an agreement is in review, waiting on the counterparty or signed. Signed terms are extracted into the repository so Contract Review Agents can use past accepted positions when suggesting redlines on future deals.

    Sources

  • Report names Mercor among US data vendors selling RL environments to Chinese labs

    News
    NewsBluesky

    A ChinaTalk investigation, covered by The Decoder on Nov 11, says at least nine US data firms, naming Mercor alongside Surge AI, AfterQuery and Turing, sell roughly $500 million of AI training data to Chinese AI labs. Part of that is reinforcement-learning environments, simulated task worlds used for post-training, and the report says Chinese buyers seek the same products sold to US frontier labs to save compute. The authors argue dual-use environments in areas like coding and biology should fall under existing export controls, while saying the trade does not erase the US lead. The findings rest partly on anonymous sources and subcontractors who disclosed client names, so this should be read as a reported claim rather than a confirmed set of deals.

    Sources

  • Modal adds an Events tab to inspect execs, tunnels and HTTP requests in Sandboxes

    Launch
    Other

    Modal shipped a Sandbox Events view in its dashboard on Nov 12 that shows exec calls, TCP tunnels and HTTP requests made with Connect Tokens for an individual sandbox. Modal says about 95% of sandboxes use at least one of those three activity types. The company says it used the view internally to find long-idle sandboxes that could be shut down sooner and to debug dropped connections and slow execs. Users reach it by opening a sandbox in the web UI and choosing the Events tab.

    Sources

  • Arena ranks Gemini 4 Hydrogen in Agent and Code Arena, Inkling 2 Small tops US open models

    Other
    X

    Arena posted day-one placements for Google DeepMind's Gemini 4 Hydrogen (High): #5 overall in Agent Arena across 6,000+ real sessions, #4 on knowledge-work tasks, and the fastest model in the Agent Arena top 10. In Code Arena WebDev it debuted at #10 with 1703 points, which Arena framed as a cost-efficient option. Arena also said Thinking Machines' newly open-weighted Inkling 2 Small is the top US open model in Code Arena WebDev at 1647 points and in Agent Arena. Replies on X were mostly neutral, with some users questioning Hydrogen's lower Text Arena score.

    Sources

    1 more source
  • Nine serverless degradation incidents hit GLM 5.3 and DeepSeek V4.1 Flash on Fireworks

    Complaint
    OtherXBluesky

    Fireworks' status page logged nine separate degraded-performance incidents on Nov 12 PT for serverless models. Five involved GLM 5.3 or GLM 5.3 Fast in the late afternoon and evening, three hit DeepSeek V4.1 Flash 1017 around midday, and one more hit GLM 5.3 Fast in the morning. Each was marked resolved, with individual incidents lasting from a couple of minutes to under an hour. Outage-tracking bots relayed every alert across X and Bluesky, adding to a negative day already dominated by breach news.

    Sources

  • Gemini 4 Hydrogen lands in Devin at 38.2% FrontierCode for $2.40 a task

    Launch
    XNews

    Cognition made Google's Gemini 4 Hydrogen available in Devin Desktop and Devin CLI, announcing it on X on Nov 12. On Cognition's own FrontierCode 2.0 benchmark the model scored 38.2%, within a point of GPT-6.1 Sol and just under Gemini 4 Carbon's 40%. Cognition highlighted cost, at about $2.40 per task or roughly 40% below Gemini 4 Carbon. Some observers on X noted FrontierCode is Cognition's own benchmark run inside each maker's harness, so comparisons should be read with that in mind.

    Sources

  • Perplexity Personal CFO reaches 17 European countries via Plaid bank links

    Launch
    NewsBluesky

    Perplexity's Personal CFO feature became available on Nov 12 in 17 European countries, including France, Germany, Spain, Italy, Ireland, Poland and the Netherlands. Users connect bank accounts through Plaid and ask plain-language questions about spending, subscriptions and upcoming balances, and the assistant can build budget trackers or spending charts that refresh as new transactions arrive. Access is read-only, so the agent cannot move money, and in the EU Plaid provides the link as a regulated account information service. French tech press framed it as Perplexity following Anthropic into personal finance, while discussion on Bluesky was limited.

    Sources