Anthropic AI Updates: September 15, 2026

1. Anthropic’s CI Job Volume Grew 25x in Six Months and Broke Its Test Selection Service

Anthropic. The engineering team published an account of rebuilding test impact analysis after agentic coding overwhelmed it. Engineers ship 8x more code per quarter with Claude authoring 80% of it, and test coverage grew 10x, driving a 25x increase in CI jobs over six months. Three temporary patches bought 70 days, 29 days, and less than a day respectively before the service needed a redesign. Source

2. The Fix Was Replacing a Singleton Writer With a Journal and a Rollup Consumer

Anthropic. The redesign moved test selection from a monolithic singleton to a stateless, horizontally scalable system backed by an in-memory data store. Instead of one writer maintaining all test history, multiple listener workers append results to a journal while a separate consumer process rolls up historical data, removing the memory bottleneck. One engineer completed the work in three weeks. The post recommends teams designing for AI-native development assume 25x load growth within two quarters. Source

3. Claude for Financial Advisors Shipped With Eleven New Wealth Platform Connectors

Anthropic. The product bundles connectors and workflow skills for meeting prep, portfolio analysis, compliance review, and post-meeting documentation, delivered as a plugin with guided setup. Eleven new connectors launched alongside it, including Addepar, BlackRock, Charles Schwab, Envestnet, iCapital, Orion, SS&C Black Diamond, Wealthbox, Wealth.com, Vanguard, and Zocks, on top of existing Microsoft 365, Salesforce, and DocuSign integrations. Anthropic recommends Enterprise plans for registered investment advisers because of audit log support, and all client-facing output requires advisor review before delivery. Source

4. Claude Code 2.1.271 Added Per-Command Domain Allowlists and Fast Mode in Remote Sessions

Anthropic. The release adds allowed_domains on individual Bash, PowerShell, and Monitor calls in auto mode with sandboxing, so the hosts a command needs are reviewed with it and opened only for that command. Fast mode became available in Claude Code Remote sessions on both cloud and self-hosted runners. Other additions include omitClaudeMd in agent frontmatter to run subagents without user, project, and local CLAUDE.md files, --accept-command <sha256> for pinning plugin installs to a previously displayed command, and a modelPricing multiplier up to 10x for internal chargeback rates. Source