A house newsletter · daily · researched and written by the Kurzia AI
Thursday 5 November 2026 · Issue 4 · 4 min read
Goldman pulls coding agents from bankers; Nooks takes $500M; Meta strands Conductor users
The story that matters most
Bloomberg · 3 days ago
Goldman pulls AI coding tools from its bankers
Takeaway: guidance updated in October
Goldman Sachs has pulled Anthropic's Claude Code and Microsoft's GitHub from its traders and bankers, leaving them to engineers and quantitative 'strats', under internal guidance updated in October, reported by Bloomberg on Nov 2. As recently as July the bank's CIO was publicly praising prototypes built by non-engineers, so this is an abrupt reversal of a year-long push, not a further rollout.
The stated reasons are productivity and risk, and the risk that bites is agents running on employee endpoints with their owners' credentials. Replit's CEO says similar restrictions are spreading across finance, and Replit's own advisory now tells companies to run agents in controlled cloud environments, not on laptops. Vendors selling 'anyone can build software' into banks lose; vendors of governed agent environments gain. Goldman is also lead underwriter on Anthropic's IPO while cutting its own staff off Anthropic's flagship tool.
Rounds and exits
Nooks — Nooks — AI revenue agents for sales teams; growth capital, non-dilutive; $500 million from General Catalyst's Customer Value Fund; valuation not disclosed.
Laza Medical — Laza Medical — AI-powered robotic imaging for cardiac cath labs; Series B; up to $160 million; led by GE HealthCare; valuation not disclosed.
Sentra — Sentra — data security governing what AI agents can reach; Series C (per SecurityWeek); $150 million led by Capital One Ventures; over $260 million raised to date.
Phaze — Phaze — remote computers for people and AI agents, from the Parsec founders; seed; $8.75 million; led by a16z; MCP server for computer-use agents.
Shackleton AI — Shackleton AI — agentic operating system for construction; seed; $5 million; led by Battery Ventures; over $1 billion of project value managed since June.
Sequoia Capital — Sequoia raised about $10 billion in new funds and is closing its London office, its Arc accelerator and its Scouts program, per its memo (reported by TechCrunch).
Where the money went. Per Crunchbase, seven billionaire-led companies led by OpenAI and Anthropic took more than half of US seed-to-growth funding this year, over $260 billion in ten months.
What the AI labs and big platforms did
Kimi K3.1 Pro — Moonshot AI put the weights of Kimi K3.1 Pro on Hugging Face: a multimodal agentic model with a 1-million-token context window, 2.8T total parameters and 104B active per token, served through an API on platform.kimi.ai. Open weights at this class squeeze the long-context and agentic premiums closed labs charge, and make self-hosting a frontier agent realistic for buyers with cost or data-control constraints. The model card benchmarks it against Claude Opus 5.5 and GPT-6 Astra on coding and agent suites, and lists vLLM and SGLang as recommended engines. Hugging Face
For your portfolio
Conductor · yesterday
Conductor
Takeaway: shuts down on Nov 20
Conductor, a Mac coding-agent app launched in July 2025, said on Nov 4 it is joining Meta. New sign-ups stopped that day; Conductor Cloud goes read-only on Nov 13 and shuts down on Nov 20, when user data deletion begins, with refunds for active subscribers. The local app keeps new models and critical fixes until early 2027.
That is the real risk of the coding-agent tail: acqui-hire consolidation at the labs strands customers on two weeks' notice, and the buyer owes them nothing. Teams that reorganized engineering around a sub-scale tool now have a migration project, not a refund. Ask the founders you back: which AI tools does the team depend on that a lab could buy next month — and what is the fallback if it does?
Hacker News · yesterday
Gemini
Takeaway: kept up to three years
A post on Hacker News (606 points) walks through Google's own support pages: on personal Google AI subscriptions, saved Gemini chats train models whenever Keep Activity is on, a subset is read by human reviewers, and reviewed chats are kept up to three years — surviving deletion of your activity. The only opt-out is turning Keep Activity off, which drops chat history after 72 hours.
Paying for a consumer subscription buys none of the data handling enterprise plans promise. Any company whose staff paste client material into personal Gemini accounts — or any consumer AI account — has a data policy gap, not a vendor problem. Question for your founders: does the data policy cover the AI accounts staff pay for themselves, or only the tools IT procured?