Your AI agents are about to get a team
Agent tools are getting easier to copy and cheaper to run, so the question is where your margin comes from. Rakazo charges no seat fee, which means you pay your model and computer providers directly and should price with that cost in mind. Frevana is betting that agents belong to the whole team, not one person's chat window. If you build agents, the defensible part is the workflow and the people who depend on it.
- Frevana Multiplayer lets teams share agents
- Rakazo: a free, open source Grok Bot rival
- Google's AMD-powered private AI chip
- Quick hits from around the AI world
- Today's deals and the Lounge AI Index
Top stories
All news →What happened: Frevana launched Multiplayer on November 15, 2026, letting a team and its AI agents work together in one shared workspace.
- Team members join through an invite link, and users choose whether their own credentials travel with anything they share.
- Edits to a shared skill reach the whole team, and each agent shows a status the team can read, with a person able to take over.
- Ready-made apps include Ad Clone, AI Product Video, Google Ads and AEO Team, and the homepage says agents connect to Shopify, Amazon and Google Ads.
Why it matters: Agents that teams can share and manage look more like shared software than personal tools, so expect buyers to ask for this. If you build agent apps, check that you can share them without exposing a buyer's login details.
What happened: Rakazo, built by Elie Steinbock, is an open source platform for always-on AI bots, positioned as an alternative to xAI's Grok Bot.
- Each bot keeps its own memory, scheduled routines and conversation history after the chat closes, and gets its own computer with browser, terminal and file access.
- Owners can connect OpenAI, Anthropic, Google, OpenRouter or Vercel AI Gateway, or run local servers such as Ollama, LM Studio, llama.cpp or MLX.
- The software has no seat fee, and the listing shows 380 upvotes and 45 comments on Product Hunt.
Why it matters: With no seat fee, your costs depend on model and compute usage, so run the numbers before you pick a platform. Support for local models also lets you test which model to bet on without locking into one vendor.
What happened: Google Cloud added the TPU 8tx to its eighth-generation TPU lineup on November 9, 2026, using AMD processors to keep private AI data protected.
- The 8tx uses the silicon of the TPU 8t training chip, but swaps the Arm-based Axion host for fifth-generation AMD EPYC Turin processors.
- AMD says prompts, model weights and intermediate data stay encrypted in host memory, in transit and while the TPU runs.
- Google claims up to 2x lower latency than the 8t with a near-zero performance penalty, though CDO Trends notes the comparison was against the training chip, not the 8i.
Why it matters: If your customers handle sensitive data, encryption during model runs is a sales point you can now point to on Google's platform. Check which chip fits your workload, since Google compared this one against its training chip, not its inference chip.
Quick hits
- Talyn runs Claude Code and Codex in cloud sandboxes on plans you already pay for, free with 10 box-hours a month. Product Hunt →
- Raschka explains how language models track word order, from GPT-2's lookup tables to rotary position embedding. magazine.sebastianraschka.com →
- HuntMonitor shows a live Product Hunt race, with snapshots every 15 minutes and sorting by points, comments and speed. Product Hunt →
- One writer argues AI-written code should be open source. AI Lounge's full brief is being written. jakub.jankiewicz.org →
- Steve Blank looks at what comes next for Nvidia. AI Lounge's full brief is being written. Read →


