Codex reboots a server and keeps going; Meta buys Conductor; Uber counts four agent rides

🔥What's hot

🛠New tools & products

• Otari — is an open-source gateway that every request from your apps and agents passes through, so you set allowed models, whether web search and code execution are on, and spending limits in one place. It answers the two questions every client steering committee asks — what can this thing spend, and how hard is it to leave a provider — without touching application code. The gateway is Apache 2.0 with a hosted option at otari.ai; budgets sit on users and application keys and are checked before each request. Mozilla AI

  “Moving to a different model or tool provider should take about as much effort as changing a config value.”

• Claude Marketplace — is Anthropic's catalogue of partner agents and products that plug into Claude, newly expanded across security, developer infrastructure, creative and productivity tools. The commercial mechanic is the story: eligible enterprise customers can put part of the money they already committed to Anthropic toward these third-party products, which quietly makes Anthropic the procurement channel. There are now over 25 partner agents and products listed; one of them, Chainguard Harness, builds an app from source inside an isolated microVM and then probes the running app for exploitable holes. Anthropic

• Seventeen working bots on one laptop and one subscription. An operations manager at a Korean B2B distributor published the full bill for seventeen Claude Code bots that file customer complaints, send statements, watch Slack for errors and price products: US$100 a month for a Claude Max plan, nothing on the API, running on a MacBook he already owned. It is the cheapest counterweight you have to a client who thinks agent automation starts with a platform purchase and a cloud budget. Fourteen of the seventeen never call a model at all — they are ordinary scripts shuttling data into Slack — and his own upkeep has fallen to roughly an hour a week, against about ten in the first couple of months. Bots in Production

  “US$100 a month, one laptop that doesn't sleep, and about an hour a week of my time. The bill was never the hard part.”

• Conductor — is a Mac app for running several coding agents side by side in separate workspaces, and it is being acquired by Meta, with the hosted side wound down fast. If anyone on your teams runs it, this is a migration with a date on it; more broadly it is Meta buying its way into the coding-agent tooling layer rather than building it. Sign-ups stopped on 4 November, cloud workspaces go read-only on 13 November, the cloud shuts down on 20 November with data deletion beginning, and the local app keeps getting new models and critical fixes until early 2027. Conductor

💬What people are saying

📰News

• Uber — told analysts on its Q3 call that during its busiest week of October only four rides were ordered by an AI agent, alongside 500 Uber Eats orders; CFO Balaji Krishnamurthy put total volume that week at around 320 million trips and orders. It is the cleanest number you can hand a client who thinks the agentic shopping wave has already arrived. CEO Dara Khosrowshahi noted booking a ride in the app takes about 50 seconds, so an agent saves nobody much time. TechCrunch

  “It is very, very early and very limited traction that we are seeing so far”

• HumanLayer — which makes approval and collaboration tooling for coding agents and published one of the first widely copied skill packs in August 2025, is joining Factory to work on software-factory tooling for large engineering teams. The second tooling startup absorbed in a day: the practices layer around agents is consolidating into a few enterprise harness vendors. The existing platform stays up for a transition of at least 30 days, HumanLayer will be open sourced, and paying customers are being refunded for the last several months. HumanLayer

• Kimi K3.1 Pro — is Moonshot AI's new open-weight model for agent work, handling text and images, with weights downloadable from Hugging Face and an API that speaks the OpenAI and Anthropic formats. Open weights that drop into existing tooling are what clients mean when they ask about leaving a single vendor, and this one is aimed squarely at agentic coding. It carries a 1-million-token context window and is a mixture-of-experts design with 2.8 trillion parameters total but only 104 billion active per token. Hugging Face

• Ripple-65B Preview — is a new open-weight model for agentic coding from Ricursive Intelligence, which says the whole thing — pretraining experiments, data work, evaluation and final runs — cost $32M, around a tenth of what it reckons the previous open-weight leader cost. If that holds, the price of a credible coding model keeps falling, which changes the build-versus-licence conversation inside bigger clients. Secondary coverage lists an Apache 2.0 licence and a Terminal-Bench score of 80.5 as claims that are not independently confirmed. Ricursive Intelligence

• SWE-2 Flash — is Cognition's new small coding model, available now in Devin Desktop and the Devin CLI at one-fifth the price and three times the speed of its larger sibling SWE-2. The interesting move is positioning rather than capability: the pitch is that everyday jobs like code review, documentation updates and automations no longer need the expensive model. The weights are not public; Cognition says it is building a programme to share them with trusted industry partners and research labs, citing cybersecurity misuse risk. Cognition

• Haiou — is a jailbreak bundle for OpenAI Codex that rewrites the agent's trusted configuration so it slips in advertisements for a third-party AI gateway, disclosed by Check Point Research, which calls the technique prompt adware. It shows the agent's own config files are now a thing attackers edit for money, which belongs in any review of what developers are allowed to install. Check Point Research named the phenomenon prompt adware; Haiou is Chinese for seagull. The Hacker News

• US voters back bans and constraints on data centres in more than 30 localities across nine states. In Tuesday's elections, residents of more than 30 localities across nine states backed bans, moratoriums or new constraints on data centres, Semafor reports — Edgewater, Florida wrote a ban into its city charter, and Perry, Ohio attached financial and operational conditions to hyperscalers seeking tax breaks. Local politics is becoming a delivery risk for the capacity your clients are quietly assuming will exist, and lenders are already treating opposition as a credit question. Tech companies are trying to relabel the sites as 'compute campuses', with mixed success. Semafor

💡One takeaway