Vercel Labs open-sources agent-desktop, accessibility-tree computer use for any agent
Vercel Labs released agent-desktop 1.0, an Apache-2.0 Rust CLI and C library that lets coding agents operate native apps on macOS, Windows and Linux. Instead of screenshots and pixel clicks, it reads the OS accessibility tree and hands the agent JSON snapshots with stable element refs. Actions run headless by default, so agents can work in parallel without taking over the user's cursor. The builder reports large token savings on dense Electron apps, though those figures are self-reported, and the launch drew thousands of likes and broad developer reposting.
Model-agnostic desktop automation for agents could replace costly screenshot-based computer use in internal tooling and QA.
Trojanized Claude Code extension in Cursor's marketplace hit up to 5,000 users
Cursor disclosed that malicious versions 2.1.312 and 2.1.313 of Anthropic's Claude Code extension were live in its VS Code Extension Marketplace for about 11 hours on Oct 29, roughly 2:27am to 1:35pm PT, for Windows and macOS. Up to 5,000 users may be affected. The tampered extension collected host details, added an attacker SSH key, and set up persistence for a reverse SSH tunnel, although Cursor says the tunnel step usually failed without admin rights and the attacker server has been taken offline. Anthropic's official extension was not compromised. Reuters, WIRED, BleepingComputer, Microsoft and several security firms covered the incident, and many developers on X and Bluesky warned others to check their installs.
Check dev machines for Claude Code 2.1.312/2.1.313 via Cursor's IOCs, remove rogue authorized_keys, and tighten extension allowlists.
Abridge launches Assistant, a chart-aware clinical AI helper for care teams
Abridge launched the Abridge Assistant, which CEO Shiv Rao pitched on X as the AI assistant for healthcare and said is already in use at more than 300 health systems. Per the product page, it answers clinical questions using the patient's chart with cited sources, summarizes what changed for upcoming patients, pre-fills calculators such as MELD and CHA2DS2-VASc from chart data, and drafts handoffs, referral and work notes for clinician review; licensed content includes American Heart Association material. A team member said five document types shipped at launch with a path to validate more. Clinicians and investors on X were broadly enthusiastic, while some replies stressed that trust and showing what evidence each agent used will decide adoption.
Ambient-scribe vendors are expanding into agentic clinical decision support inside the EHR; health-tech teams should watch scope creep and governance.
Sierra says it passed $400M ARR, adding over $150M in Q3 alone
In an Oct 30 post, Sierra co-founders Bret Taylor and Clay Bavor said the customer-experience agent company crossed $400M in annual recurring revenue about two and a half years after launch, with more than $150M of that added in Q3. They said growth is speeding up: doubling from $100M to $200M took six months and from $200M to $400M took five. The post credits customers pushing agents into longer, revenue-driving interactions, faster builds with its Ghostwriter tool (one large US retailer went live in 27 days), outcome-based pricing, and work with nearly half the Fortune 50. On X the milestone was widely relayed and mostly read as proof that customer-service agents are where AI revenue is landing first, with a few skeptics asking about resolution quality and calling the approach old-school enterprise CX.
Strong evidence enterprises are buying outcome-priced support agents at scale; benchmark build-vs-buy for CX automation against Sierra.
Replit details Agent Habitat security model as Cloud Agents push its 'self-driving company' pitch
Replit published an engineering post on Oct 28 laying out the security architecture behind its Agent Habitat, the governed workspace in Replit for Business where agents reach company systems and data under scoped permissions and approvals. The post says Replit's existing isolation blocked last week's disclosed kernel flaws and that it patched tooling and pushed firewall rules centrally, and it frames least-privilege access and monitoring as core principles. The New Stack tied this to CEO Amjad Masad's claim that Fortune 500 firms will run as agent-driven companies by 2030, and to the new Cloud Agents that let non-engineers change a codebase and learn from code-review feedback via shared skills. Chinese tech outlets covered the architecture paper, and Bluesky reaction was mildly positive while X replies were split with critics recalling the 2025 database-deletion incident.
Shows how a major agent platform scopes agent permissions and governance; useful benchmark before letting agents touch internal systems.
You.com and Fireworks pair up to give open models keyless live web search via MCP
You.com announced a partnership with Fireworks AI that lets open models served on Fireworks, such as Kimi, DeepSeek, GLM, Qwen, MiniMax and gpt-oss, pull live, cited web results through You.com's MCP server. A You.com guide published Oct 30 shows Fireworks' OpenAI-compatible Responses API calling the MCP server server-side, with a free keyless profile capped at about 100 queries a day, plus examples for tool-calling agents and LangChain. Fireworks' own Responses API docs describe the server-side MCP execution that makes this work. Reaction on X was small but positive, with developers liking that agents can start without an API key.
Teams on open models can add grounded web search to agents with one MCP tool entry instead of building a search pipeline; check the free tier's query cap.
Gemini 4 Carbon takes first place on Mercor's APEX-Agents benchmark
Mercor said Google's new Gemini 4 Carbon debuted at the top of its APEX-Agents leaderboard with 84.2% Pass@1 and a 90.4% mean score, the only model above 90% mean. Gemini 4 Argon came second, ahead of Claude Sonnet 5.5. Carbon led the management consulting and investment banking tracks, while Argon still led corporate law. Mercor's post was widely reshared, including by Google DeepMind leaders, which keeps raising APEX's profile as a benchmark for professional-services agent work.
APEX-Agents is becoming a reference for picking models for long, multi-app knowledge work; the Gemini 4 models now lead it.
Legora launches Ontological Legal Research in the US
Legora launched Ontological Legal Research in the United States on Oct 29, pairing comprehensive case law and statute data with a proprietary ontology that maps how authorities connect, rank and get overruled. The company says the aim is to stop AI from citing outdated law, misjudging authority or missing exceptions, and that where publishers won't partner it gathers primary law itself, including by physical scanning. The launch builds on a September announcement of an AI-native citator. Law.com's weekly legal tech rundown led with the feature on Oct 30, and Legora's CEO pitched it on stage at its Precedent NYC event.
A concrete example of grounding an agent with a domain knowledge graph plus owned data to fight stale or wrong citations, not just better prompting.
LangChain open-sources Content Studio, its in-house Slack design agent built on Deep Agents
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.
Having agents write assets as code plus sandbox-per-thread is a reusable blueprint for internal creative tooling.
Runway launches App Studio, which builds no-code creative apps from a plain-language description
Runway introduced App Studio, which turns a plain-language description into a custom creative tool with its own interface that generates images, video and audio using Runway's models plus a team's Brand Kits and Workflows. Apps can be used privately, shared with a team, or published for anyone with a Runway account, with enterprise controls on top. Runway pitches it as a way to turn one power user's prompts and settings into software the whole team can use. It was Runway's most-viewed post of the week, and follow-ups added Character Swap for replacing characters or locations in video and turning short films into games.
Creative teams can now ship their own internal AI tools without engineering, which shifts who builds and maintains them.
Harvey extends Legal Agent Bench to test end-to-end brief writing on US case law
Harvey's research team released LAB: United States Case Law, an extension of its Legal Agent Bench that has frontier models research and draft full litigation briefs from real public case records sourced via Free Law Project. Model-written briefs are compared head-to-head, and Harvey reports that a small group of frontier models can now produce complete first-draft briefs in one shot without attorney guidance, though a long tail of failure modes remains. Harvey says it will make this research more accessible through a partnership with Grubel. On X, a Harvey researcher and legal-tech accounts shared the results positively, treating it as a meaningful step for long-horizon legal agents.
A new long-horizon agent eval built on real case records; useful for teams benchmarking models on complex, multi-step drafting tasks.
Cognition acquires Corridor to build out Devin Security
Cognition is bringing in Corridor, a startup that built a security layer for AI-native engineering teams, and its team will work on Devin Security. Corridor's co-founders announced the deal on Cognition's blog on Oct 28, saying they will work with Cognition's president, CTO and CSO and framing the next 6 to 12 months as a critical window as AI agents become capable of full attack chains. The move follows Cognition's Devin Security 2.0 upgrade two days earlier, which added frontier-model code scanning and automatic fix PRs. On X, posts described it as folding real-time guardrails for coding agents into Devin, with mostly neutral, informational reactions.
Security for AI-written code is consolidating into coding-agent vendors; teams on Devin should evaluate Devin Security versus standalone AppSec tools.
Perplexity releases pplx-embed-v2-context embeddings trained without gold-chunk labels
Perplexity published pplx-embed-v2-context, a family of contextual embedding models in 9B and 0.6B sizes, with weights on its Hugging Face organization and an accompanying arXiv paper. The models are trained by distilling graded chunk-level relevance from the query-aware context compression model used in Perplexity's answer pipeline, so retrieval can surface both answer chunks and the evidence needed to verify them. The paper introduces two benchmarks, Context-bench and Q2D-Chunk (released as a public leaderboard), and claims state-of-the-art results on them and on ConTEB, plus strong visual document retrieval. Retrieval researchers shared it on X, though discussion volume was modest.
Teams running RAG can test open contextual embeddings that retrieve supporting context, not just a single best chunk, with one vector per chunk.
Glean lets agent creators import and export agents and speeds up its agent engine
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.
Portable agents and a large MCP catalog reduce lock-in; the default AI notice is an EU AI Act compliance cue.
Nvidia reportedly weighs credit guarantees for Nscale's Anthropic data center
Nvidia is reported to be in talks to provide billions of dollars in credit guarantees so Nscale can raise debt for a roughly 460MW West Virginia campus that Anthropic plans to use under a ~$45B compute lease. The Information broke the story and Reuters, Bloomberg and TechCrunch followed; neither company has confirmed a deal. The structure discussed is a residual-value style backstop covering construction and power costs. Reaction on X and Bluesky was largely skeptical, with many posts calling it circular financing since Nvidia would be supplier, investor and guarantor. The news lands while Nscale pursues a planned IPO.
Capacity leased from neoclouds like Nscale may depend on vendor-backed debt, so watch delivery risk before committing GPU workloads.