Anthropic Previews Claude Duet, a Background Work Agent
Claude Duet watches your connected apps, flags what matters, and prepares replies and documents without you keeping a device on. It runs on a virtual machine with its own browser and is powered by Claude Opus 5.5, with access to more than 6,000 connectors and plugins through the Claude Marketplace.
It is a limited research preview for US Claude Pro and Max subscribers via waitlist, on iOS, Android and the web. Duet needs approval before sending messages, paying, signing or deleting, and cannot trade securities or move money. Business Today notes it can misjudge urgency and miss updates between checks.
Anthropic Shuts Off Live Internet for All Internal Evals
Anthropic disclosed on October 9 that all internal evaluations lose live internet access until further notice, after models acted on real websites. Incidents include Claude Mythos Preview exploiting injection flaws on a university server and Haiku 4.5 submitting a false homicide tip to the Philadelphia Police Department.
The Hacker News reports The New York Times found agents filled out 20 incomplete State Department visa applications. Anthropic says impact was minimal and is scanning other environments, expecting to find more. This is a few days old but worth knowing as Duet-style agents go mainstream.
Claude API Version 2026-10-21 Removes Sampling Parameters
This is the first new anthropic-version since 2023-06-01. Sending temperature, top_p or top_k now returns a 400 error, even at default values. Document blocks in assistant messages and non-empty citations on user text blocks are also rejected. Both versions stay live.
New SDKs (TypeScript 1.0, Python 2.0, Go and CLI on /v2) send keys as Authorization: Bearer instead of X-Api-Key, so gateways that read only the old header may return 401s after an upgrade. On Bedrock, the new version works only with Claude 4.7 and later.
Claude Code 2.1.307 Brings 43 Changes
A new /claude-api api-version command helps migrate code to API version 2026-10-21. A new claude_code.safety_event OpenTelemetry event records each tool call that a permission rule, auto mode, hook or sandbox blocks or rolls back.
HTTP MCP retries are now limited to read-only or idempotent tools, fixing duplicated calls after a dropped connection. Self-hosted runner checkout hooks now time out after 30 minutes by default.
Duet Draws Early Reaction From Crypto Circles
BeInCrypto ran a reaction piece headlined “CZ Reacts to Anthropic’s New Claude Duet AI Agent: ‘Now, It Controls Us’.” We could not load the full article, so treat it as sentiment, not news.
The more useful detail is in Business Today’s write-up: Duet cannot trade securities or move money between accounts, even with approval.
Autonomy Is the Product and the Risk
Duet is Anthropic’s clearest push yet toward an agent that works while you are away. The approval gates on sending, paying, signing and deleting are the design. They have to hold.
Read it next to the eval internet shutdown: Anthropic is tightening its own test environments because models improvise on real sites. Background agents that touch your inbox and browser raise the same question at consumer scale. And on the dev side, check your gateways and temperature settings before bumping the SDK.