Highlights
What happened at the companies we cover, by day, newest first. Every item links to its sources.
Showing 1–17 of 17 highlights
Wednesday, Nov 4 17 stories
Dreaming keeps project memory and AGENTS.md notes tidy on paid plans
LaunchNewsHN
On November 4 Lovable introduced Dreaming, a background process on paid plans that checks project memory and AGENTS.md rules against code and chat, then rewrites or removes stale notes without changing app code. Dreaming does not consume credits; AGENTS.md edits appear in project chat and version history. The same day Lovable published an engineering blog on how the company repeatedly rebuilt its agent stack, citing 70 million projects and sandbox forking.
Sources
Baseten Router public preview: coding-task routing within 0.8 pts of Opus at 72% lower cost
LaunchNewsOtherX
On November 4 Baseten introduced Baseten Router as part of Baseten Grid in public preview. The router chooses between frontier and cheaper open models per coding task and reported pass rates within 0.8 points of Opus 5.5 at about 72% lower cost on DeepSWE, SWE Atlas QnA, and Terminal-Bench, beating random routing with statistical significance among commercial routers tested. Opendoor’s eng lead cited auto-routing as a way to stop overpaying for simple work.
Sources
AI Gateway adds more ZDR providers for Kimi K3.1 Pro
LaunchNews
Vercel expanded Zero Data Retention routing for Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed on AI Gateway beyond Moonshot alone. Teams can require ZDR per request via provider options or enforce it for every gateway call in the dashboard, and requests fail closed if no ZDR provider is available. Both models keep a 1M-token context with image input, with Highspeed aimed at interactive coding agents. The changelog walked through AI SDK, Chat Completions, and coding-agent setup paths.
Sources
Perplexity launches Teammates for ongoing Computer work
LaunchNewsX
Perplexity launched Teammates, always-on agents that own long-running projects inside Perplexity Computer with tools, skills, memory, and Slack collaboration. Teammates can monitor work proactively, hand tasks between agents, and present results for review, initially for Max subscribers on web with iOS coming soon. The launch reframes Computer from single-session tasks toward delegated business workflows across marketing, sales, support, and ops.
Sources
Harvey embeds Jus Mundi Jus AI for arbitration research inside Assistant and Workflows
LaunchNewsOther
On November 4 Harvey and Jus Mundi announced an agent-to-agent partnership that puts Jus AI inside Harvey Assistant, Agents, Vault, and Workflows for international arbitration teams. Research stays citation-backed against awards, treaties, and institutional materials while using matter context already in Harvey, under Harvey security controls. Joint customers can activate at no extra Harvey cost by year-end if they also have a Jus AI plan.
Sources
Suno lets Pro/Premier users start songs from MIDI files
LaunchNewsXOther
Suno shipped MIDI uploads on Create so Pro and Premier users can drag .mid/.midi files onto the page and prompt how closely the model should follow the notes. MIDI carries exact pitch and timing without audio, giving composers precise control compared with remixing WAV/MP3 uploads. Official guidance covers exact playback, motif development, and re-instrumentation for scores and student projects. Coverage on November 5 amplified the launch across producer communities. Studio also continued adding finer pitch and tempo-aware import controls around the same window.
Sources
Parloa and Deloitte Canada form alliance and reseller pact for agentic CX
NewsNews
On November 4 Parloa announced an alliance and reseller agreement with Deloitte Canada to deploy its AI Agent Management Platform for Canadian contact-center modernization. The pitch pairs Parloa agents with Deloitte’s integration, governance, and change-management services, and highlights Parloa Zones for Canadian data residency. This is a go-to-market expansion for North American enterprise CX rather than a new product SKU.
Sources
Profound runs Zero Click Singapore for 400+ marketers
NewsNewsXOther
Profound hosted Zero Click Singapore 2026 at Raffles on November 3 with more than 400 marketing leaders, then published a November 4 recap. Speakers from Singtel, Standard Chartered, Unilever, DHL Express, and L’Oréal discussed AEO measurement, Marketing Engineer roles, and agentic shopping journeys. Profound’s research team shared citation and commercial-intent findings from tens of millions of AI-search citations. The company used the event to introduce itself ahead of opening a Singapore office later in the year. Sessions were released on demand as Zero Click Singapore Unlocked.
Sources
Kimi K3.1 Pro lands on Modal Shared Endpoints with Day-0 SGLang bringup
LaunchXOther
On November 4 Modal announced Kimi K3.1 Pro from Moonshot is available on Modal Shared Endpoints with token billing and an OpenAI-compatible API. The 3T-class MoE model targets coding, financial research, and long-horizon agents with text/image input and 1M context. Modal said it co-validated checkpoints with the SGLang project for Day-0 serving. This expands the shared-endpoint catalog for teams that want open-weight agent models without standing up dedicated GPUs first.
Sources
Together Compiler continually optimizes open-model GPU inference
LaunchNewsOther
Together AI launched Together Compiler on November 4, describing an agent-driven continual compilation system that rewrites GPU inference code for open-source models. The company claims up to about 2.6x higher throughput within a day of model release and says the system already serves large production token volumes, with a limited beta. For eng leaders buying serverless or dedicated inference, this is a direct pitch against static kernel stacks when new open models ship weekly.
Sources
Replit opens Tokyo office, deepens SMFL Asia enterprise push
NewsNews
Replit announced its first Asia office in Tokyo on November 4, following London in September, and deepened its partnership with Sumitomo Mitsui Finance and Leasing (SMFL). At a joint press conference, Replit said SMFL used the platform so one employee finished in a week work that previously took six people six months, and both sides plan to expand enterprise app building across Asia. The move ties to Replit’s Self-Driving Company pitch for agents with security and controls.
Sources
Replit rewrites sandbox engine on microVMs for agent isolation
LaunchOtherNewsHN
Replit published Part 1 of its new sandbox engine write-up on November 4, describing a full rewrite onto Cloud Hypervisor/KVM microVMs so each sandbox gets its own Linux kernel instead of a shared Docker host kernel. The company says acquiring a sandbox takes roughly 500 ms with warm pools, and it migrated hundreds of millions of projects without data loss. A self-audit of 6,211 Linux-CNA CVEs from June through November 4 claims 96.1% avoided via minimal guest kernels and early patches—critical for eng leaders running untrusted agent code.
Sources
Community roundup: stay on 2.46.4 stable; try 2.47 for Agents and queue mode
NewsNews
Independent coverage on November 4 summarized n8n 2.47.0 for operators: npm stable/latest still pointed at 2.46.4 while beta/next/rc tracked 2.47.0. Highlights called out the Assistant kill switch, Agent version history with independent publish, Postgres client SSL cert fields, fairer queue-worker job distribution, and Migration Report improvements for the upcoming 3.0 cut. Upgrade tests on a small Docker instance reported clean migrations with Code nodes still working.
Sources
OpenEvidence launches free Learn medical education platform
LaunchNewsX
OpenEvidence launched Learn, a free personalized medical education platform for trainees and clinicians. It includes Test Prep with over 50,000 board-style questions spanning USMLE through specialty boards, a Socratic AI Preceptor for case-based tutoring, spaced-repetition flashcards, and CME credit for questions clinicians already ask. Johns Hopkins and Penn Med are early feedback partners. The launch extends OpenEvidence from point-of-care answers into career-long training on the same evidence base.
Sources
Everlaw ships MCP connector to Legora for eDiscovery workflows
LaunchNewsBlueskyX
Everlaw's November 4 product release added connectors so users can connect workflows to Harvey and Legora. Legora availability covers the US, UK, EU, Canada, and Australia, while Harvey's MCP connector is early access in the US only. Everlaw told customers to confirm data-flow requirements with their Legora rep before enabling the link. The integration lets litigation teams keep review in Everlaw while handing work into Legora's agentic legal environment.
Sources
Devin adds hierarchical budgets for enterprise ACU spend control
LaunchNews
Cognition introduced hierarchical budgets so admins can allocate Devin Agent Compute Units across business units, teams, and projects up to three levels deep, with separate cloud and desktop controls. When a budget is exhausted mid-task, the session stops; engineers get alerts and can request more capacity, with optional efficiency-score-based auto-approval. The New Stack coverage ties the feature to customers scaling Devin beyond single-team pilots into org-wide deployments.
Sources
Cognition launches SWE-2 Flash: 3x faster, ~1/5 the cost of SWE-2
LaunchNewsOther
Cognition released SWE-2 Flash on November 4, a 276B-total / 12B-active coding model post-trained from Thinking Machines' Inkling-Small with a new mid-training stage before RL. The company says Flash runs about three times faster than SWE-2 at roughly one-fifth the price, and is available in Devin Desktop and CLI, free for paid users through November 30. Weights stay private for now while Cognition builds a secure partner-sharing program, citing cybersecurity dual-use risk after raising ExploitBench from 11.3% to 34.8%.
Sources