Anthropic AI Updates: October 2, 2026

1. Claude Code Mods Let TypeScript Functions Rewrite Prompts, Tool Calls, and UI

Anthropic. Anthropic introduced mods for Claude Code, small TypeScript functions that hook into events Claude Code emits and can run before, after, or around them. A mod can rewrite a prompt before it reaches the model, block, rewrite, or retry tool calls, approve or deny permission requests, redact secrets from tool results, and edit, replace, or extend the interface in the terminal or desktop app. Mods ship inside plugins, install through /plugin, and work in the CLI and desktop app; /diff is now itself a built-in mod that can be disabled or replaced, and a sec-default mod loads first on managed enterprise deployments to enforce permission policy. Mods run unsandboxed with full machine access, so Anthropic advises installing them only from trusted sources, and enterprises can restrict them through managed settings. Source

2. Barclays Targets Claude Code for Half Its Developers by the End of 2026

Anthropic. Barclays is expanding its use of Claude across operations, client service, and software engineering. Its retrieval-augmented Colleague Knowledge Assistant, live since 2025, is used by more than 16,000 employees supporting over 20 million UK retail customers and has handled more than 1 million searches, while Claude models in Global Markets classify, enrich, and route about 120,000 emails a day. The bank aims to put Claude Code in the hands of 50% of its developers by the end of 2026 and a majority of its software engineers by 2027, with legacy system modernization and code quality among the stated goals. Source

3. A Physicist Used Claude on 400 Candidate Problems and Produced 36 Manuscripts Across 18 Fields

Anthropic. In a guest post on Anthropic’s research site, physicist Matthew Schwartz argues that researchers should look for “Claude-shaped problems” that fit current models’ breadth, coding, and math skills, while humans supply the conceptual direction. Over three months he and 19 coauthors screened about 400 candidate problems and produced 36 manuscripts across 18 fields, using an open-source harness called BootLoops that first computed 15 previously unknown elliptic Feynman integrals. Other results include converting 4,452 economics replication packages from proprietary to open-source code, analyzing 5.7 billion mutation pairs for gene conversion evidence, and assembling a word-stress database covering 6,072 languages. Source

4. Claude Code 2.1.287 Adds a “You Should Know” Side Agent and MCP URL Prompts

Claude Code 2.1.287, tagged October 1, shipped the mods runtime alongside a built-in mod called “You should know,” in which a side agent watches the session and flags things the user or Claude might miss (enabled with /plugin enable cc-plugin-you-should-know@builtin, first-party sessions with telemetry on). MCP servers on the 2025-11-25 protocol can now send URL prompts, for example to sign in; servers that stop connecting after the update need "bareElicitationCapability": true in their config. The agents view gained an n:<text> filter for session names and tasks, OpenTelemetry user_prompt events gained a prompt_text field, and self-hosted runners using Anthropic-managed git get a built-in REST-only gh api when the GitHub CLI is not installed. Source