Highlights
What happened at the companies we cover, by day, newest first. Every item links to its sources.
Showing 1–20 of 41 highlights
Wednesday, Nov 11 41 stories
Parloa deepens OpenAI tie-up with early model access and joint enterprise sales
NewsNewsXBluesky
Parloa and OpenAI announced on Nov 11 an expanded partnership under which their engineering teams will hold joint on-site sessions to improve agent response time, reliability and cost. Parloa is joining OpenAI's Alpha Model Testing Program to evaluate new models for customer service before public release, and the two plan to sell jointly to enterprises with a North America focus. Parloa also plans to list its platform on the OpenAI Marketplace, with no date given, and says Codex is used across its engineering team. Coverage on X came mostly from CX trade outlets and startup news accounts, with little debate.
Sources
n8n's Nod agent draws wide day-two interest, with praise for its candor on safety and cost
LaunchXBlueskyHN
A day after n8n Labs released Nod as a research preview on Nov. 10, the free Apache 2.0 personal agent that runs in a local sandbox with its own browser spread widely on X and Bluesky and landed on Product Hunt. n8n's launch post says Nod asks before sending, paying, posting or deleting anything not requested, keeps logins encrypted out of the model's view, and works with the user's own AI provider, but admits the permission check is itself a model that can err. The post also notes early testers found tasks costlier than expected, prompting per-task receipts, monthly caps and a check-in above $5. Builders mostly praised the approve-before-acting design and the split between agent-written drafts and deterministic n8n workflows, while some cautioned that 98% completion came from simulated test sites.
Sources
2 more sources
Together AI researchers open-source FOSCAR kernels for 2- and 4-bit KV-cache decoding
OtherNewsOther
A new preprint from Together AI and University of Sydney researchers introduces FOSCAR, an open-source kernel library that speeds up decode attention over 2-bit and 4-bit quantized KV caches. Instead of dequantizing every cached value, it folds the scale and offset into the other operand or the output of the matrix products, cutting those operations by up to two orders of magnitude while keeping exact results. The authors report geo-mean speedups on H100 of about 3.4x (INT2) and 2.8x (INT4) over FlashInfer BF16, with smaller gains on B200, and up to 341% higher peak throughput than 16-bit baselines in SGLang long-context serving. The code is published as a kernel library on GitHub, with Hopper and Blackwell-specific pipelines.
Sources
LangChain 1.5 drops Python 3.10 and adds model-router middleware and decision models
LaunchBlueskyOther
LangChain released langchain-core 1.8.0 and langchain 1.5.0 on Nov. 11, making Python 3.11 the minimum now that 3.10 is end-of-life; projects still on 3.10 will keep resolving the 1.7.x and 1.4.x lines. The core release adds decision models, and langchain 1.5.0 adds model router and auto mode middleware, with ModelRouterMiddleware marked beta. langchain-typesafe 0.1.0 shipped the same day, reimplementing its TypeSafeClassifier as a decision model and deprecating its experimental middleware. The Python 3.10 cutoff was flagged by release-tracking accounts on Bluesky, with little debate so far.
Sources
1 more source
Builders report multi-week autonomous Devin runs on SWE-2
PraiseX
Several builders on X described long autonomous Devin sessions on Nov 11. An accounting-agent startup said one Devin thread building and testing its tax agent ran for 20 days, and a mobile team said Devin on SWE-2 worked for 17 days to get a React Native app store-ready. Others praised SWE-2's speed and showed tools built with Devin in a day, and Sonar shipped a SonarQube plugin for Devin CLI to check generated code before PRs. A popular counterpoint post lumped Devin in with past AI hype that faded, so sentiment was not uniform.
Sources
2 more sources
Devin's Nov 11 release adds self-set reminders, Design Mode edits and scan triage
LaunchOther
Cognition's Nov 11 Devin release notes add the ability for Devin to schedule reminders that wake a session later, including sessions started by automations or service users. Design Mode now opens a compact prompt bar for per-property edits and lets users annotate a page screenshot before sending it to Devin. Devin Security gains finding-to-threat linking, dismissal reasons, severity letters, auto-scan schedules and clearer errors when scans fail. Other changes include an option to require a click before loading external content in sessions, better stacked-PR visibility and guided Perforce setup.
Sources
Devin lets paid users bring their ChatGPT plan and caps Teams at 100 users
NewsNewsX
Devin's pricing page now lists a use-your-ChatGPT-plan option on the Pro, Max and Teams tiers, and the Teams plan is capped at up to 100 users. PricingSaaS, which compares daily captures of the page, says the cap fell from 200 between its Nov 10 and Nov 11 snapshots, while Teams gained multiplayer sessions, SSO, three-tier cost controls and Slack channel support. The current page shows Teams at $80 a month plus $40 per full developer seat, with Pro at $20 and Max at $200. Social reaction was light, though a few X users weighed Devin's plan value against rival coding tools.
Sources
Fireworks' zero data retention turns out to be an opt-in enterprise policy
DiscussionNewsX
A VentureBeat analysis published November 11 pieced together why Notion transcripts were exposed despite Fireworks advertising zero data retention by default. Fireworks' docs say only live inference requests are held in memory, while stored responses, batch jobs, datasets and training keep customer content, and the policy that blocks storage is enterprise-only and off by default. The article notes that CEO Lin Qiao's post excluded batch inference from the products with no detected impact, though neither company has said which feature held Notion's data. On X, practitioners argued that buyers cannot see the chain of inference and cloud providers behind their SaaS vendors, and some said it is a reason to run models in-house.
Sources
Fireworks breach fallout widens as Notion revokes exposed API keys found in chats
NewsNewsXBluesky
On November 11 Notion updated its advisory to say it had revoked Notion API keys that it found in affected users' messages, which were exposed through the Fireworks AI breach. Notion has stopped serving open-weight models through Fireworks, disabled its Insights feature, and is telling affected customers to treat the transcripts as exposed and rotate any credentials they pasted. BleepingComputer recapped that the attacker reached Fireworks' development environment through a third-party cloud provider on October 23 and got billing data, model weights and datasets for some customers, following a separate secrets leak in early October. The story spread widely on X and Bluesky, where security accounts framed it as a supply-chain risk hidden behind SaaS vendors.
Sources
Profound positions itself as ChatGPT's first-party brand analytics raise threat questions
DiscussionXOther
After OpenAI opened a beta Business Hub that shows verified businesses how often ChatGPT references their site, recorded clicks and discovery topics, Profound published a Nov. 11 post calling it the first ground-truth data from an answer engine. Profound argued the Hub only counts website-linked mentions in one engine, while its platform covers brand mentions without links, per-prompt rankings, competitor share of voice, cited sources, sentiment and multiple engines, and said it will help customers pull Hub data into Profound. Skeptics argued the Hub undercuts paid visibility trackers that estimate presence by replaying prompts, with one X post saying it takes aim at tools like Profound and Peec AI. The debate is split between seeing the Hub as validation of the category and as a direct competitive threat.
Sources
Gemini Managed Agents can now execute tools inside Modal sandboxes
LaunchNewsX
Google added self-hosted execution environments to Gemini Managed Agents on Nov 11, letting the agent harness run code in a developer-controlled sandbox, and Modal is one of seven launch providers alongside Vercel, Daytona, Fly.io, Google Cloud, Cloudflare and E2B. Gemini still handles planning and orchestration, while shell and file tools run in the chosen sandbox, with sandbox compute billed by the provider. Google's docs link a Modal provider guide in the Gemini API CLI repo. On X, a developer who worked on the integration with the DeepMind team promoted running Gemini agents across many Modal sandboxes, drawing modest engagement.
Sources
Modal joins gVisor's maintainers as Google hands the sandbox runtime to CNCF
NewsXOther
Modal said on Nov 11 that it is one of three companies, with Ant Group and Tines, committing to maintain gVisor as Google donates the project to the CNCF at KubeCon. gVisor is Modal's core container runtime, and the company says it launches upward of 40 million gVisor containers a day across GPU inference, multi-node training, coding agents and sandboxes. Its post lists upstreamed work such as nvproxy GPU support, a new rdmaproxy for multi-node training at about 99% of native NCCL bandwidth, and checkpoint/restore used for memory snapshots. Next on its roadmap are multi-GPU memory snapshots, GPUDirect Storage and MIG support, and Modal is hiring kernel and runtime engineers. The X announcement was Modal's most-engaged post of the day.
Sources
Nscale reportedly weighs pausing its NYSE listing until next year
NewsNewsXBluesky
The Telegraph reported on November 10, and Reuters carried the story, that Nscale executives are weighing whether to pause their roughly $35 billion New York IPO until next year after the tech-stock sell-off. Nscale filed for the NYSE listing under the ticker NSCL in September, seeking about $3.5 billion, including $2 billion from Nvidia. The report came days after Reuters said Microsoft had declined an option to rent more capacity at Nscale's Texas site. On November 11 the story spread widely on X and Bluesky in several languages, where commenters debated Nscale's reliance on Microsoft and whether its contracted backlog holds up.
Sources
Decagon ships Alto, an in-house model that runs the talk layer of its duplex voice agent
LaunchXNews
Decagon Labs introduced Alto on Nov. 11, a conversational model that decides on each turn whether its duplex voice agent should answer, wait for confirmation, or keep the caller engaged while a separate reasoning model performs lookups and actions. Decagon says Alto averages about 700 ms responses and lifted live-call resolution by more than five percentage points without other agent changes, building on its earlier Chord speech model. It was built from a small open-weight model, fine-tuned on simulated calls, refined with reinforcement learning, tested on over 90 million simulated calls and shadowed on production traffic. Decagon also let visitors call the agent from its website for the first time; replies on X were enthusiastic, with developers asking how interruptions affect in-flight actions.
Sources
2 more sources
Vercel AI Gateway adds Google's Gemini Decision 1 and a Decisions API endpoint
LaunchBlueskyOther
On November 11 Vercel added Google's Gemini Decision 1 to AI Gateway as google/gemini-decision-1. It is a multimodal decision model that returns Boolean probabilities, choices and scores without generating output tokens, and Vercel pitches it for routing requests, policy checks, filtering agent tools and ranking. Developers can call it through AI SDK 7's experimental decide function, a TypeSafe-compatible endpoint, or a new Decisions API on the gateway; PDF, audio and video inputs only work through the AI SDK. Decision calls appear in gateway logs and count against budgets like other model requests. Changelog feeds and AI news trackers on Bluesky picked up the release.
Sources
Harvey licenses Danish legislation and preparatory works from Djøf Forlag
NewsNewsX
Harvey announced on Nov 11 a partnership with Danish legal publisher Djøf Forlag to bring consolidated acts, executive orders, amendments and preparatory works into its platform. Lawyers holding a Jurabibliotek license will be able to use the content in Assistant, Agents, Vault, Workflows and the Word add-in next to their own documents, with cited answers linking statute, regulation and legislative intent. Harvey says the content goes live at the start of next year at no extra charge, following a similar data deal with Gyldendal Rettsdata in Norway. Harvey shared the news on X, where it drew light engagement and little debate.
Sources
OpenEvidence launches free Access Intelligence for copay help and prior auths
LaunchXNews
OpenEvidence launched Access Intelligence on Nov. 11, a free set of tools inside its clinical AI that answers eligibility questions for more than 2,500 patient support programs, with program data refreshed daily. Clinicians can enroll patients or download prefilled forms, send copay offers through Patient TakeHome, and draft prior-authorization requests, appeals and medical-necessity letters grounded in guidelines with checklists and ICD-10 codes. The company says clinicians had already written over 500,000 prior-auth documents on the platform before the launch. Physicians on X largely cheered it, while a few questioned how a free tool is funded and whether health-system integrations survive once costs are surfaced.
Sources
2 more sources
Lawyers on X argue Legora and Harvey are wrappers that may lag frontier models
DiscussionX
The most-liked Legora posts of the day came from a BigLaw-focused account arguing that Harvey and Legora could hurt the firms relying on them by leaving blind spots, because wrapper products may not keep pace with frontier model capabilities. Other posters added that firms claiming to be covered just by buying Harvey or Legora are taking a risk, and that many associates quietly use general chatbots on the side despite firm platforms. A smaller set of replies defended the platforms, with one poster calling Legora's in-house dogfooding a reason to be bullish. The debate was confined to X and stayed split between platform skeptics and backers.
Sources
2 more sources
Legora and DiliTrust link contract and entity records via an MCP integration
LaunchNewsXBluesky
Legora and DiliTrust announced a partnership on Nov 11 that connects DiliTrust's contract and entity management data to Legora through a Model Context Protocol integration. In-house teams can ask about a contract's terms, search contracts in bulk for due diligence, draft from approved DiliTrust templates, and query entity details like shareholders and governance. The connector is read-only, so DiliTrust stays the system of record and its permissions apply, and it launches as an early-access beta for mutual customers. Legal tech outlets and Legora's X account shared the news with little debate.
Sources
Legora makes Google Workspace integration generally available to its Agent
LaunchX
Legora made its Google Workspace integration generally available on Nov 11, letting the Legora Agent search Google Drive, pull related Gmail threads and attachments, and check or book Google Calendar slots. Admins pick which of the three services to enable for the organization, and each user then links their own Google account. Access follows the user's existing Google permissions, and sending email or changing files and events requires the user's approval. Legora's announcement on X drew modest engagement and no notable pushback.
Sources