Skip to content

Highlights

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

Showing 1–16 of 16 highlights

Wednesday, Oct 28 16 stories

  • DeepLearning.AI and LangChain launch a free 'Loop Engineering with Deep Agents' course

    Launch
    XNews

    DeepLearning.AI released a free short course built with LangChain on 'loop engineering': designing the loops around an agent, not just the agent itself. Over eight lessons it covers the basic agent loop, a verification loop where an evaluator checks output against criteria and retries, event-driven loops that trigger runs automatically, and hill-climbing loops that improve the harness from traces. It includes code examples and graded labs. The announcement was well shared, and LangChain promoted it again on Oct 30.

    Sources

  • LangChain says over 70% of Managed Deep Agents runs now start from schedules, not humans

    Discussion
    XNews

    LangChain argued that most proactive or 'always-on' agents now shipping are really scheduled runs with a prompt, an old operating-system idea applied to agents. It published a guide to the Schedules SDK in Managed Deep Agents, which lets developers define cron schedules in code or let the agent create and delete its own schedules through governed tools. The company says that since dynamic scheduling launched in September, over 70 percent of Managed Deep Agents runs on LangSmith come from schedules rather than someone typing a message. The post was widely bookmarked.

    Sources

  • Instinct can now check out with Shop Pay at any online store, not just Shopify

    Launch
    XOther

    Instinct expanded its Shopify partnership so the agent can pay with Shop Pay at any online store, including non-Shopify merchants, using the card and address already saved in Shop. Users still approve every purchase. Shopify's help center says Instinct and Claude Duet pay anywhere using one-time payment credentials, while Meta's Muse is limited to Shopify stores.

    Sources

  • Runway Dev ships a Codex plugin so the coding agent can build Runway's API into apps

    Launch
    X

    Runway released a plugin that gives OpenAI's Codex what it needs to integrate the Runway API into an application. Developers install it once, describe what their app should generate, and let Codex write the integration on its own. It follows Runway's earlier plugin that brought its generation tools into chat assistants, and it drew strong engagement for a developer announcement. The move reflects API vendors packaging their docs and tooling for coding agents rather than only for human developers.

    Sources

  • ElevenLabs details Cue, the forced-alignment model timing every Scribe transcript

    Launch
    NewsBluesky

    ElevenLabs published details on Cue, a single forced-alignment model that places each character of a transcript on the audio timeline and now sits behind every Scribe request, subtitles, dubbing and lyric timing. It handles recordings of any length in one pass instead of stitching chunks, and the company says it aligns an hour of audio in about half a second. On ElevenLabs' own real-world benchmark across 29 languages it reported lower timing error than the open-source aligners it tested, including on noisy audio and songs. Cue is also offered directly through the Forced Alignment API.

    Sources

  • Crusoe says its Spark modular AI factories will support NVIDIA's new Vera Rubin DSX POD

    News
    X

    At NVIDIA's GTC event in Berlin, Jensen Huang unveiled the Vera Rubin DSX POD, a prefabricated design he said would change how data centers are installed and built. Crusoe quickly said its Spark modular AI factories, which it has built and run since 2020, will support the new POD. The statement ties Crusoe's recent Spark Liquid launch to NVIDIA's next rack-scale platform. It positions modular, factory-built capacity as a faster path to deploying next-generation GPUs than traditional large campus builds.

    Sources

  • Crusoe executive says every GPU arriving through 2028 is already sold, as neocloud debt costs climb

    Discussion
    News

    Speaking at a tech conference in Israel, the head of Crusoe's Israeli operations said every GPU scheduled to arrive by 2028 is already sold to customers. He also admitted that raising debt has become increasingly difficult and predicted consolidation among neoclouds that simply buy GPUs and resell capacity. The next day CoreWeave's 2031 bond hit a record yield of about 15.8 percent, and The Inference framed the two as the stage and the credit market pricing the same AI demand curve differently. The piece also noted that Crusoe's recent $700M facility is secured by its own power-generation equipment.

    Sources

  • Harvey launches an Intake artifact inside Anthropic's Claude for Legal

    Launch
    Other

    Harvey was a launch partner for Anthropic's latest Claude for Legal release, shipping a Harvey Intake artifact that business users run inside Claude. It connects to Harvey through Harvey's MCP server: routine legal questions are answered from Legal's playbooks and prior work, requests needing judgment are routed to a lawyer in Harvey, and every request is logged in Intake with assignment rules and status tracking returned to the requester in Claude.

    Sources

  • Perplexity's new voice mode runs several Computer sessions at once on OpenAI's gpt-live-1

    Launch
    X

    Perplexity updated its voice mode so a single spoken session can juggle multiple Computer agent sessions, start new tasks, and create and manage Projects. The company says it runs on OpenAI's gpt-live-1 real-time model. It is available on the Mac app and the web for Pro and Max subscribers. The update pushes voice from a question-answer feature toward a way of supervising several running agents, and it drew one of the larger engagement counts among Perplexity's posts this week.

    Sources

  • Paper Instruments RL-tunes GLM-5.3 Flash on Fireworks into a cheap knowledge-work model

    News
    Other

    Fireworks published a case study on Paper Instruments, a five-person team that used the Fireworks Training API to run reinforcement learning on GLM-5.3 Flash with rubric-graded synthetic tasks for memos, analysis and slides. The resulting Ultramarine 1 Flash is now served on Fireworks serverless and its weights are public. On the startup's own internal benchmark it edged out a leading closed model at roughly 2% of the per-task inference cost, though that result is self-reported. The team says the training run took about two weeks and the whole effort cost around $300K, with Fireworks handling GPU tuning and trainer/rollout numerical alignment.

    Sources

  • Replit releases ViBench 2.0, testing agents that build whole products feature by feature

    Launch
    XHNOther

    Replit's AI team published ViBench 2.0, a long-horizon coding benchmark where an agent starts from an empty project and adds up to 18 features to products modeled on apps like Uber or Google Docs. It spans 17 products, 241 feature requests and 544 browser-run test plans, graded by an agent whose accuracy Replit checked with planted bugs. Top frontier models fully pass only about half to two-thirds of builds, and most failures happen where features interact. A cheap open-weight model passed most individual plans at a fraction of the cost, and the benchmark joins Fireworks' Specialized Intelligence Index.

    Sources

  • AI Gateway adds virtual models to change agent model settings centrally

    Launch
    News

    Vercel AI Gateway now supports virtual models, saved configurations that apps and agents call through a custom model ID. Teams can swap the underlying model, reasoning effort, caching or fast mode in the gateway without redeploying code. The feature also covers provider choice, fallback models and requirements such as zero data retention. Vercel positions it for steering coding agents that do not expose those settings themselves and for attributing spend.

    Sources

  • Vercel CI enters beta, running tagged Turborepo tasks on Vercel

    Launch
    News

    Vercel opened a beta of Vercel CI, which runs any Turborepo task tagged for CI on its own infrastructure rather than only builds. Setup is a single tag in turbo.json, and Vercel suggests letting a coding agent configure it through a published skill. The company says its own repos saw push-to-green times drop by two to six times. It pitches the speedup as necessary for CI to keep pace with code written by agents.

    Sources

  • Eleven medical societies license their flagship journals into OpenEvidence via Wiley

    News
    News

    Wiley said eleven independent medical societies, spanning neurology, epilepsy, psychiatry, haematology, geriatrics and obstetric imaging, have made their flagship journals available inside OpenEvidence. The content sits alongside Wiley's existing license to OpenEvidence, which since March includes the Cochrane Database of Systematic Reviews and hundreds of Wiley journals. OpenEvidence's chief medical officer framed the move as giving physicians traceable, specialist-reviewed citations they can open directly. Wiley said it plans to add more societies and extend the arrangement to health systems outside the US.

    Sources

  • Cognition reportedly weighs new round at $70B-$80B, weeks after $48B Series E

    News
    NewsX

    Business Insider reported that Cognition has received investor offers valuing it at $70 billion to $80 billion and is seriously considering raising before year end, per people familiar with the matter. The talks are described as early and the terms could change; no lead investor or round size was named. The report comes only weeks after the September Series E of more than $2 billion at a $48 billion valuation, when the company said run-rate revenue was nearing $900 million. This remains a report, not a confirmed raise.

    Sources

  • Genspark opens its internal agent sandbox infra to developers as SB-Box beta

    Launch
    XOther

    Genspark released SB-Box in beta, exposing the VM sandbox layer that powers its own agents to outside developers via the gsk CLI and a REST API. Each sandbox is a full virtual machine with its own kernel, restored from a snapshot rather than booted, with a claimed sub-second median start. Sandboxes auto-freeze when idle, keep memory and disk, resume on the next request, and can be cloned or rolled back by version. Genspark pitches it as a code interpreter for coding agents via a single SKILL.md file, as a persistent machine for long-running agents, and as forkable environments for agentic RL. The launch post drew modest traction, with about 60 likes and two dozen replies.

    Sources