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 20 highlights

Tuesday, Oct 27 20 stories

  • Arena adds Connectors to Agent Mode for Gmail, Drive, Linear, Vercel, Supabase and more

    Launch
    OtherX

    Arena (formerly LMArena) shipped Connectors in Agent Mode, letting users plug everyday tools such as Gmail, Google Drive and Calendar, Linear, Vercel and Supabase straight into agent sessions instead of copy-pasting between tabs. This lets the community run longer, real-world multi-step tasks end to end, and those tasks feed the Agent Arena rankings.

    Sources

  • Synthesia ships a Localization API for 20-language dubbing and Brand Kits in its Syren video agent

    Launch
    News

    Synthesia shipped two updates aimed at enterprises using it at scale. A new Localization API lets teams dub a video into up to 20 languages in one request straight from their CMS, LMS or content pipeline, track each language until it's ready, add languages later, and control access with dedicated API key scopes. Separately, Enterprise customers can now pick a workspace Brand Kit in Syren, the prompt-to-video agent launched earlier this month, so the first draft already uses the company's logos, colors, fonts and avatars. Both extend a month of steady Syren and API updates.

    Sources

  • LangSmith now tracks prompt-cache hit rates per agent and per thread

    Launch
    X

    LangSmith added tracking for how often an agent's LLM calls hit the provider's prompt cache. Teams can see an agent's overall hit rate as one number, break it down thread by thread, and chart it over time. Since cached calls are cheaper and faster, LangChain frames this as a direct lever on agent cost and latency. The announcement drew more replies than most of the company's posts that week.

    Sources

  • LangChain open-sources Content Studio, its in-house Slack design agent built on Deep Agents

    Launch
    XNews

    LangChain published a long write-up on Content Studio, its internal design agent that makes on-brand videos, social graphics, decks and diagrams from a Slack prompt, and released the full harness as an open-source repo. Every asset is written as code (HTML slides, chart specs, animation timelines), so a small edit is a single value change rather than a full regeneration. Each Slack thread gets its own sandbox, long jobs run on async subagents, and the thread shows the plan and asks for approval before anything ships. The team says it took four months of feedback to teach the agent its taste. The launch post was LangChain's most-viewed of the week.

    Sources

  • Instinct opens iMessage group chats to all users, including friends without accounts

    Other
    X

    Instinct made its group-chat mode available to every iMessage user. Once added to a thread, a group-scoped Instinct works for everyone in it, including people who have not signed up, and only knows what is shared in that chat. Each member's personal Instinct still asks before sharing anything with the group. The feature had been in limited release since early October.

    Sources

  • Runway Ads goes GA for every paid plan, generating ad variants from live performance data

    Launch
    XNews

    Runway made Runway Ads available on every paid plan. The agent connects to a team's ad accounts, starts from ads that are already performing, and generates on-brand variants that people approve before they are published across platforms. Runway says its own marketing team used it to grow weekly ad volume from under 100 to about 1,000 with the same headcount while cutting cost per subscriber by 41 percent. It named Lovable among the teams already using it.

    Sources

  • Together details evidence-gated auto repair for its GPU clusters, now generally available

    Launch
    Other

    Together AI published a deep dive on auto repair for Together GPU Clusters, with passive health checks, repair recommendations and auto-remediation policy now generally available. Detection runs both inside the customer VM (Xid errors, DCGM ECC and thermal signals, node problem detector) and on the host (InfiniBand link state, reachability). Every repair needs a failed health check behind it, and customers choose per fault or per Xid code which fixes run unattended and how long to wait for jobs to go idle. Fixes escalate from reboot to reprovision to host migration, protected by cooldowns, concurrency budgets and a circuit breaker for correlated failures. Letting customers bring their own detectors via node labels is in preview.

    Sources

  • Gamma adds Airtable and Clay connectors so its agent builds decks from live data

    Other
    X

    Gamma announced two more inbound connectors for its Gamma 5 agent: Airtable on Oct 26, turning bases like roadmaps, pipelines or content calendars into presentations, and Clay on Oct 27, pulling enriched GTM data and lead lists into personalized pitch decks, account plans and QBRs. They extend the connector set (Notion, Slack, Granola, HubSpot, Linear) the agent can reference as context while generating or editing.

    Sources

  • Legora connects to the iManage MCP Server for permission-aware firm knowledge

    Launch
    Other

    Legora and iManage said they are connecting Legora to the iManage MCP Server, building on an existing API integration for browsing and saving documents. Through MCP, Legora workflows will be able to move documents and update their metadata inside iManage, and firms using Insight+ can pull institutional knowledge from past matters into Legora. iManage permissions decide what each user can retrieve, and iManage stays the single source of truth. Both companies pitch it as safer, governed access to firm work product from an AI agent.

    Sources

  • Stoel Rives rolls out Legora firm-wide after a 100-lawyer pilot

    News
    Other

    Stoel Rives, an Am Law 200 firm with about 350 attorneys in ten offices, picked Legora as its firm-wide AI platform. The decision came after a multi-office pilot with more than 100 lawyers and a formal security review run by the firm's own AI and cybersecurity practice. Litigation teams used it to map case facts across large document sets, draft early filings and speed up matter intake. Research staff said it returned fewer but better-targeted results than competing tools. The deal adds to the string of big-law adoptions Legora has announced this month.

    Sources

  • Lovable adds Events & Hooks so companies can watch and block workspace actions

    Launch
    News

    Lovable added Events & Hooks for Business and Enterprise workspaces. Events notify a company's HTTPS endpoint when a project is created, published, moved to another workspace or deleted. Hooks go further and let that endpoint allow or block an action before it happens, for example scanning a chat message for sensitive data before Lovable processes it. The same day it added a Scrive connector for e-signatures and national eID checks and renamed Knowledge to Instructions, continuing a month of governance features aimed at IT teams.

    Sources

  • Glean lets agent creators import and export agents and speeds up its agent engine

    Launch
    News

    Glean's October 27 release lets any agent creator, not just admins, import and export auto and workflow agents to move them between deployments and teams. Autonomous agents now run on an improved execution engine by default that finishes multi-tool workflows faster. The MCP server library gained 57 remote templates, and Google's Gemini 4 Argon joined the model hub, with Glean sharing its own eval comparison. Assistant will also show an AI-generated response notice by default from November 10, aimed at EU AI Act transparency.

    Sources

  • n8n ships 3.0 beta focused on reliability and security, with breaking changes

    Launch
    OtherXBluesky

    n8n released the 3.0.0 beta on October 27 and plans a stable 3.x release in early December. The major version bundles breaking changes: manual runs always go to workers in queue mode, in-memory binary storage is removed, the default Code node timeout drops to one minute, and self-hosting is now Docker-only with npm/npx installs dropped. Legacy nodes such as Function, Cron, Interval and the old OpenAI node are removed, and safer defaults include always-on encryption key rotation and unverified community packages turned off. A built-in Migration Report flags affected workflows and auto-fixes many of them. The 2.x line gets 12 months of security fixes after 3.x goes stable, and support for 1.x ends December 31, 2026.

    Sources

  • Baseten adds four engineering leaders from Meta, Anyscale, Google and NexHealth

    Hiring
    News

    Baseten announced four senior engineering hires to scale its inference business. A former Meta PyTorch inference acceleration lead becomes Head of Inference Platform, and Anyscale's former Head of Engineering joins to lead forward-deployed engineering with customers. A former Kindo CTO and longtime Google Drive architect joins the inference stack team, and a former NexHealth engineering head takes over Product Platform. The company said it keeps hiring across inference, model performance, infrastructure and forward-deployed roles.

    Sources

  • Modal backs General Catalyst fellowship with up to $25K compute per undergrad

    Other
    X

    Modal partnered with General Catalyst on a research fellowship for about 20 undergraduate builders. Each fellow can receive up to $25K in Modal compute to pursue ambitious projects without shrinking them to fit their hardware. Modal's announcement was one of its most-engaged posts of the week, with over 200 likes. The program seeds Modal usage among early-career AI researchers and future hires.

    Sources

  • Modal Sandboxes now snapshot their filesystem automatically on any exit

    Launch
    Other

    Modal released Exit Snapshots in beta for Sandboxes, capturing filesystem state whenever a sandbox ends. That covers clean terminations and entrypoint completion as well as idle timeouts, lifetime limits and out-of-memory kills. Developers no longer need to snapshot proactively to keep work from a sandbox that died. Modal pitches it for resuming coding agents after an idle timeout or salvaging results from unexpected crashes.

    Sources

  • Clay makes a post-trained model router the default for its Claygent research agent

    Launch
    OtherX

    Clay introduced auto-mode as the new default for Claygent, routing each row to a cheaper or stronger model depending on how hard the research task looks. The router is an open-weight model post-trained with Applied Compute and served on that partner's platform, making decisions in under a second at hundreds of millions of queries a month. Clay says auto-mode cost 47% less on production traffic than sending everything to the most capable model, with per-category quality floors to avoid degrading tasks like contact lookups. Billing follows whichever model ran each row, so costs vary within a column, and existing Claygents keep their pinned model. GTM builders on X noted the change the same week.

    Sources

  • Report: EliseAI is quietly building an AI-first property management system for 2027

    News
    News

    A Thesis Driven exclusive, based on interviews with multifamily executives who have seen demos, reports that EliseAI is developing its own AI-native property management system. The platform is said to be targeted at institutional operators and slated for an early 2027 launch, which would put EliseAI in direct competition with Yardi and RealPage. The report ties the effort to RealPage's earlier move to block new EliseAI integrations. EliseAI has not announced the product publicly, so this remains a report rather than a confirmed launch. The move would shift EliseAI from an AI layer on top of incumbent systems to owning the core system of record.

    Sources

  • Abridge ships in-visit orders, chart-aware inpatient notes and an AI evaluation dashboard

    Launch
    Other

    Abridge's October release pushes its ambient AI beyond note drafting into orders, coding and governance. Orders discussed in the room are now staged in Epic for sign-off, and inpatient H&P and progress notes pull in labs, vitals and imaging changes with inline citations. Its Query Engine now raises documentation queries and lighter cues for outpatient visits, and accepted inpatient diagnoses are written straight to the chart. A new Evaluation Dashboard lets health-system AI governance teams see each metric's method, refresh cadence and known limits.

    Sources

  • Cognition releases FrontierCode 2.0, a harder mergeability benchmark where top model scores ~45%

    Launch
    NewsX

    Cognition published FrontierCode 2.0, the next version of its coding benchmark built with open-source maintainers who grade whether an agent's change would actually be merged. The new set pulls from more repositories, adds security and performance tasks, roughly triples 'thoroughness' blocking criteria, and requires larger multi-file changes. Claude Opus 5.5 leads at about 45.5%, with big drops for most models versus FrontierCode 1.1 and weak open-source results. Cognition notes models can score lower at higher reasoning effort, and X users questioned comparing models across different vendor harnesses.

    Sources