Skip to content

The Monday brief: Week of Nov 9–15, 2026

15 stories from 15 companies

Mercor raises $750M Series D at $20B to push evals and post-training into enterprises

Mercor closed a $750 million Series D that values it at $20 billion, twice the roughly $10 billion mark from its October 2025 round, with NVIDIA, Salesforce, Zoom, Benchmark, General Catalyst and others participating and an employee tender included. CEO Brendan Foody's post says the money goes into evals, post-training infrastructure and simulated work environments so enterprises can grade and tune agents, often starting from open models, on their own workflows. The Deeptune acquisition, which rebuilt hundreds of enterprise apps as agent practice environments, is central to that pitch, and the company says it now spends about three times more on inference than on salaries. On X the reaction was mostly congratulatory investor posts and news reposts, with some analysts questioning how the platform revenue turns into profit.

Signals a well-funded vendor for custom agent evals and post-training on your own workflows, a buy option versus building eval infra in-house.

Fireworks breach fallout widens as Uber driver license tally passes 7,800

Fallout from Fireworks AI's October intrusions kept spreading on Sunday, with state breach notices showing driver's license data for roughly 7,870 Uber drivers in three states was taken, and California and Illinois counts still pending. Security press also detailed Cursor's account that a key it had issued to Fireworks was exposed and used to copy shared data, which led to a trojanized Claude Code extension reaching up to 5,000 users for about 11 hours. Fireworks' own bulletin attributes the activity to a financially motivated LLMJacking group and says inference, Training SDK and Nexus, which are zero data retention by default, were not affected. Reaction on X and Bluesky is overwhelmingly negative, with security accounts framing it as an AI vendor supply-chain warning and some questioning why driver data sat with an inference provider.

Audit what data and credentials you share with inference vendors and rotate any keys issued to Fireworks.

LangChain confirms $300M Series C at $3B and repositions around LangSmith

LangChain officially announced a $300M Series C at a $3B valuation on Nov 13, led by CapitalG with GIC, DFJ Growth, IVP and Sapphire Ventures. TechCrunch notes that is more than double the $1.25B valuation from roughly a year ago. The company is centering its pitch on LangSmith as a framework-agnostic platform for running, improving and governing agents, claims more than 7,000 organizations use it, and promises a batch of new LangSmith capabilities on Nov 20. The CEO's announcement post drew over a thousand likes and reposts on X. Some replies argued the money is going to LangSmith rather than the open-source framework.

Teams on LangGraph/LangChain should expect a heavier push toward paid LangSmith runtime, eval and gateway features, with new launches due Nov 20.

Builders pick apart Poppy 1.0 a day after Sierra's agent protocol hit every Shopify store

The day after Sierra shipped version 1.0 of its Personal Agent Protocol (Poppy), developers on X and Bluesky spent Saturday unpacking how it works. Most posts focused on the mechanics Sierra described: each Shopify store now auto-publishes a Poppy file, Meta's Muse looks for it first, and customers scope agent access via OAuth on top of MCP, OpenAPI and UCP. Sierra's release also lists 60 new design partners, taking the total past 100, with Mercury, Synchrony and Turkish Airlines going live. Reaction was mostly positive about a sanctioned entry point replacing site scraping, though some replies asked how people will see and cap what an agent shares or buys, and a community Sylius plugin implementing the spec already appeared on Bluesky.

Teams building agents or commerce sites should evaluate publishing a poppy.json and OAuth scopes now that Shopify and Muse default to Poppy.

Cursor breach report spreads; developers told to check extension versions and use Privacy Mode

Cursor's Nov 13 incident update spread widely on Saturday across security press, X, Bluesky and a small HN thread. The report says a key Cursor issued to Fireworks was exposed, letting an attacker copy mostly 2025 agent traces from non-Privacy-Mode sessions, under 0.05% of users, some holding unredacted credentials. Those led to internal tools and a credential used to plant tampered Claude Code builds 2.1.312 and 2.1.313 in Cursor's marketplace cache for about 11 hours, reaching at most 5,000 users. BleepingComputer adds the malware installed persistent SSH backdoors; Cursor says it moved all inference off Fireworks and rotated credentials. Reaction was largely negative, with posts urging version checks and Privacy Mode and calling stored agent logs an attack surface.

Audit where AI vendors store your agent traces and keys, enforce secret redaction in logs, and verify editor extension provenance.

06n8n

n8n's Nod agent draws wide day-two interest, with praise for its candor on safety and cost

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.

An open-source reference for local computer-use agents with approval gates and cost caps worth studying before building your own.

Runway opens Praxis-1 robot world action model under Apache 2.0

On November 11 Runway released weights for Praxis-1, its first world action model, under Apache 2.0 with inference and fine-tuning code, including LoRA, for commercial use. The 27B mixture-of-experts diffusion transformer predicts video and robot actions together, and Runway reports 46.2% success on NVIDIA's RoboLab-120, ahead of FLUX 3 Action, Cosmos3 Nano and pi0.5. It ships on GitHub and Hugging Face with LeRobot support and vLLM-Omni integration, and runs from H100 and B200 down to RTX 5090 and Jetson Thor. Gigazine adds that Runway's own tests found GPT-6 Astra took about 15 times longer and cost about 51 times more per task. X and Bluesky reaction, much of it from Chinese and Japanese developers, was enthusiastic, with a few people noting it has only been validated on partner robots.

Robotics teams get a commercially licensed open world action model to fine-tune on their own arms, which could cut teleoperation data needs.

Vercel ships Workflow SDK 5 with per-run encryption, regional runs and big latency cuts

Vercel released Workflow SDK 5, the durable-execution library behind its Workflows product, rebuilt for long agent loops after more than 4 billion steps and 700 million runs. Every workflow and step payload is now encrypted end to end with AES-256-GCM using a per-run key, and runs can live in any Vercel Function region with failover instead of only iad1. Vercel's own benchmarks against v4 show step-to-step p75 overhead falling from 618 ms to 109 ms and a 1,020-step run dropping from 21 to 3 minutes, which Vercel stresses reflects one workload in one region. Reaction on X was broadly positive, led by Vercel's CEO, while one developer cautioned that the new run attributes are stored in plaintext.

Teams running long agent loops can get durable execution with encryption and data residency without building their own queue layer.

Cognition launches Devin Apps: private team web apps hosted on Devin's cloud machine

Cognition introduced Devin Apps, letting teams describe an internal tool and have Devin build and host it as a private web app on Devin's own cloud computer. Apps inherit the session's MCPs, integrations and codebase, get a stable sign-in-only URL shareable to people, an org or the whole enterprise, and keep prior versions serving if a new deploy fails. They run on a single machine that hibernates via memory snapshots and bills only while awake, so Cognition pitches them for dashboards, internal tools and staging copies rather than high-traffic production. Reaction on X was enthusiastic, with users praising the pace of shipping, while several asked how apps are isolated from workspace secrets and whether they can reach systems behind company auth.

Coding agents now host the internal tools they build, so plan guardrails for agent-run apps touching MCPs, secrets and internal data.

Decagon ships Alto, an in-house model that runs the talk layer of its duplex voice agent

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.

A concrete duplex pattern: a small RL-tuned talk model fronting a slower reasoning model to cut voice latency without false confirmations.

Developers debate Perplexity Decider v1.3's gains as the open-weight classifier hits HN

Perplexity's Decider v1.3 and Flash v1.3 decision models, released with a span-extracting locate mode and Apache-licensed weights, reached the Hacker News front page on Nov 13 with over 100 points. Trade coverage stressed the low per-request pricing and that all benchmark wins over TypeSafe's Jev are self-reported. Commenters were split: some found it strong and cheap on their own tests, while others said the main model barely moved versus v1.2 on a public decision index and questioned the flash-tier trend. Perplexity's CEO separately said its decision models now serve about a billion requests a month on OpenRouter.

Cheap open-weight decision models can replace LLM calls for routing, judging and PII extraction, but teams should benchmark on their own data.

ElevenLabs ships Scribe v3 and v3 Live speech-to-text with realtime diarization

ElevenLabs released two new transcription models: Scribe v3 for recorded audio and Scribe v3 Live for realtime use, both in ElevenAPI and with v3 Live also in ElevenAgents. The company reports 27% fewer word errors than Scribe v2 across 102 languages for v3, and 43% fewer than Scribe v2 Realtime for v3 Live, which also gains speaker labeling and context-aware turn detection for the first time. Developers get keyterm prompting of up to 1,000 terms, plain-English formatting rules, EU/India/Singapore data residency, and a migration that is just a model ID change. Japanese trade outlet TRAIL noted pricing was not disclosed and all accuracy figures are self-reported, and X posters echoed that they want to test it on noisy real-world audio. The official launch post drew hundreds of likes and quotes on X.

Teams running meeting notes, contact-center or voice-agent STT can trial a drop-in model swap with realtime diarization and better turn-taking.

Genspark moves Deep Research to a post-trained open Qwen3.5-397B model

Genspark now runs its Deep Research agent by default on a Qwen3.5-397B model it post-trained with Baseten Applied Research, replacing its earlier fine-tuned open model. On Genspark's own grader the new model scored 0.759 versus 0.612 for the base model and 0.738 for Opus 5.5, while crawling about five times more sources per query. The recipe paired reinforcement learning for the search phase with supervised fine-tuning for report writing, and the model serves with NVFP4 weights on two B200 GPUs per replica. Co-founder Kay Zhu said Genspark keeps the weights, evals and training code. On X, replies were mostly upbeat, though several noted the margin over Opus is small and measured only on Genspark's internal evals.

Shows a production team swapping a closed frontier model for a post-trained open model to cut cost and own the weights.

Users show Replit Agent running Google Ads on its own as approval questions surface

Two days after Replit gave its Agent the ability to run Google Ads and buy supplies through Stripe's Link and Visa, early users began posting results on X. One builder said asking Replit to grow an app led Agent to set up a Google Ads account, start campaigns and generate tens of thousands of paid impressions within 48 hours, and CEO Amjad Masad amplified it as autonomous businesses arriving. Replit's blog says Agent stops at checkout for owner approval, with approval checks enforced in code, and Replit Support repeated in replies that Agent prepares checkout for the user to review and approve. Reaction was largely enthusiastic, while one reply asked that each charge record show who approved it and for how much.

Teams letting agents spend money should copy Replit's code-enforced approval gates and audit trails before enabling ads or purchasing agents.

Hugging Face's llama.cpp team ships v1.0 of llama.cpp and ggml with stable APIs

The llama.cpp team at Hugging Face released v1.0.0 of both llama.cpp and the ggml tensor library, about three and a half years after the project began. Going forward, breaking changes to the public C API or ABI will trigger major-version bumps under explicit semantic-versioning rules, aimed at the many downstream apps that embed the libraries. The team cited support for more than 150 model architecture variants, wider packaging across Homebrew, Debian, conda-forge, Winget and Docker, and CI scaled on HF Buckets and Jobs with self-hosted hardware runners. Reaction on Hacker News, across three threads, and on X was upbeat, with commenters asking about KV-cache quantization guidance for coding and a Linux desktop app.

Teams embedding local inference get a semver-stable llama.cpp/ggml API, reducing upgrade risk for on-device and private AI products.

Start this Monday.