Harvard: agents write 30% more code, ship 12% more work; 1Password brokers agent logins; Uber puts 30,000 on ChatGPT Work

1Password for Agents — is a credential broker: the agent never receives your password, it identifies itself to the vault with its own key, and the approval happens on a device you hold before the login is filled in for it. Client pilots stall at the login screen for sites with no API, and this is the first mainstream answer worth recommending. Supported at launch are Claude Duet, ChatGPT Work and Dots, Meta Muse and Grok Bot; an agent that is hijacked after sign-in can still act inside the account. explainx.ai

OpenShell — is NVIDIA's open-source runtime for fleets of agents: each one runs in a sandbox where the kernel decides which files, system calls and network destinations are allowed, and real credentials are attached outside it. A buildable answer when a client asks what sits between an autonomous agent and production. Policy changes are formally checked and anything granting risky new access waits for a human; Apache 2.0, with 17.8k stars and SDKs for Python, TypeScript, Go and Rust. GitHub

Agent Compute Engine — is a Google Cloud service, in preview later this year, that gives each agent its own hardware-isolated microVM: a small virtual machine that starts like a container but holds no secrets. It is the infrastructure behind Google's own agents turned into a product, which makes agent isolation a line item rather than a project. When a tool call ends the host checkpoints the machine so it uses nothing until the next call; the open-source sibling for Kubernetes, Agent Substrate, handled 500 activations per second in a 200,000-agent test. The New Stack

“We started from one rule: Assume an attacker controls everything inside the sandbox, and make sure they find nothing there to use”

OpenAI CLI — is OpenAI's official command-line tool for the API: make an image, upload a file, submit a batch or inspect results from a terminal without writing an application. Worth separating in a client's head from Codex: one scripts API calls inside a pipeline, the other is the coding agent that edits your repo. Binaries ship for macOS, Linux and Windows, and a setting, OPENAI_UNTRUSTED_STDIN=1, hardens it when piping JSON from a source you do not trust. GitHub

Supabase MCP Server block — lets a Supabase app publish its own actions as agent tools, so a customer can use the product from inside ChatGPT or Claude without leaving the conversation. Permissions are the interesting part: each tool call runs as the signed-in user, so the access rules already in the database decide what an agent can reach. Installing it copies an edge function into the project with sign-in middleware and a whoami tool; the demo app let a customer list products and generate video ads from chat. Supabase

Claude API 2026-10-21 — is Anthropic's first new dated API version since 2023: the temperature, top_p and top_k sampling settings are gone, and sending any of them now returns a 400 error on every model. Worth an audit in client codebases, because wrappers, defaults and config files quietly set these in places nobody remembers. Seven SDKs and the ant command-line tool shipped new major versions on it, moved the key into an Authorization: Bearer header, and switched licensing from MIT to Apache 2.0. Tech AI Wire

Safety monitors let more than half the attacks through. Anthropic's alignment team attacked the blocking monitors inside Claude Code's Auto Mode and OpenAI's Codex Guardian using an adversarial agent that had been told how those monitors work. On average, over 55% of attacks got through. Ammunition for the next meeting where a classifier is presented as the control that makes unattended agents safe. The monitors were quiet in normal use, stopping under 1% of 4,000 harmless actions, and attacks split across context compactions or several agents remain unsolved. explainx.ai

ARTEX — an open-source autonomous penetration-testing agent, was used through October against IT and software vendors, Wiz reports, with the activity overlapping a Chinese state-aligned actor Microsoft tracks as Storm-0940. This is the concrete version of the AI-attacker question clients keep asking, now with logs rather than speculation. Password-spray bursts ran about a second apart for 20 to 60 minutes, some sessions lasted 24-30 hours, and new addresses appeared minutes after blocks; the author took the code closed-source on October 8, by which point it had 12,000 forks. Wiz

Agentic AI Foundation — the Linux Foundation's home for agent standards, took four more donations at its San Jose conference and now holds ten projects: Anthropic's Agent Skills packaging format, OpenAI's Open Responses specification for talking to any model provider, Hugging Face's OpenEnv training environments and Alibaba's OpenSandbox. Handy in lock-in conversations: the plumbing is drifting onto neutral ground while the vendors compete on models and products. The foundation opened in December 2025 with three projects: MCP, goose and AGENTS.md. explainx.ai

Claude Models on Google Cloud (Anthropic managed) — is now generally available: the complete Claude feature set reached through Google Cloud's own identity permissions and a single Google Cloud invoice, with new features arriving the same day as on Anthropic's own API. Removes a familiar procurement objection at Google-shop clients, who no longer trade feature lag for a tidy bill. Usage draws down an existing Google Cloud commitment, and Fast Mode on Opus 5.5 is generally available at $8 and $40 per million input and output tokens. Anthropic

ChatGPT Work — goes from a 1,000-person pilot to Uber's whole workforce of about 30,000, running by default on the cheaper GPT-6.1 Sol model and plugged into Gmail, Drive, Calendar and Slack. The reference deal to cite when an executive asks who has actually taken agents past the pilot stage; note that 80% of early users were not engineers. Uber's engineers stay on Codex, and the company reportedly used up its entire 2026 Claude Code budget by April. TechCrunch

DeepSeek — is finishing a round of about $15 billion, roughly 100 billion yuan, according to Reuters and Bloomberg reports the company has not confirmed; CATL and Tencent are named among the largest backers. Keeps the cheap open-weight option firmly in play for clients weighing Chinese models, whatever Washington does next. It comes two days after US federal civilian agencies were ordered to strip DeepSeek technology out, and would roughly double the original 50 billion yuan target. Trending Topics

OpenAI — has trimmed the price on its new round to $1.25 trillion, down from the $1.4 trillion it put to prospective backers earlier in October, while trying to close up to $30 billion led by Coatue and UAE-linked investors. A useful temperature check for clients: the pace-setter is reading a colder room ahead of a listing it now expects in 2027. The Nasdaq Composite fell 2.1% over the week, its worst stretch since June. Startup Fortune

Laminar — is the in-house agent harness Ramp built after trying managed agent APIs, open-source frameworks and third-party platforms, and it now runs every agent product at the company. The build-versus-buy question clients ask; Ramp has published no running costs or comparisons, so six weeks is only half the story. It keeps agent sessions separate from the sandboxes where code runs; LangChain, whose CEO praised the design, says picking a model by task difficulty cut median cost per thread from $2.61 to $0.94. The New Stack