English

Product LaunchesAnthropicClaude Code

Anthropic Enhances Claude Code with Multi-Agent Projects and Introduces API Optimization Tools

By the editors ／ Model: gemma-4-26b-a4b-it-qat-6bitAs of: 2026-09-18

Anthropic has introduced an updated Projects feature within Claude Code, enabling users to manage a team of AI agents that coordinate with each other. In this feature, multiple "threads" run different tasks in parallel, with each thread functioning as a Claude Code cloud session on its own branch. A "coordinator" directs the workflow, and if threads modify the same code, conflicts are resolved through merge conflicts similar to a standard pull request. These projects benefit from a shared memory, shared goals, and a library of files and artifacts. While threads currently run in the cloud, Anthropic stated that support for local tools and code is coming "very soon." The updated feature is currently in beta for select Claude Pro and Max subscribers, with expanded access to Pro, Max, Team, and Enterprise users expected in the coming weeks.

To improve efficiency, Anthropic also announced new tools for the Claude API designed to manage costs and performance. The /claude-api prompt-audit command identifies common prompting "anti-patterns"—such as redundant verification rituals or contradictory instructions—that can inadvertently increase costs and decrease accuracy. In testing, using this audit helped reduce costs by 14.6% and increased accuracy by 5.3% on average. Additionally, the /claude-api cost-optimize command profiles token spend and suggests reductions through prompt caching, batching, and bounding output. For iterative optimization, the /claude-api hillclimb command performs a search to find the best configuration of effort levels and model choices to maintain performance while minimizing costs.

Regarding the internal use of AI, Anthropic released metrics showing that as of August 2026, AI plays a significant role in its research and development (R&D). According to the company, 26% of its R&D tasks have reached "Level 4" automation, meaning AI can complete most tasks from a high-level prompt while humans supervise. When including tasks where AI "collaborates" with humans (Level 3), the involvement exceeds 90%. Furthermore, as of August 2026, approximately 30,000 AI agents were running simultaneously on Anthropic's primary internal platform for research and engineering work.

Sources

  1. Claude Code relaunches Projects to manage multiple AI agents in the cloud （The Verge AI, 2026-09-17）
  2. Claude Blog
4 more sourcesHide sources
  1. AIへの「念のためプロンプト」はもう逆効果　Claudeのトークンを浪費する6つのアンチパターン （ITmedia AI+, 2026-09-18）
  2. Anthropic
  3. Anthropic社内で約3万体のAIエージェントが同時稼働、AI研究開発の26％をClaudeが主導 （GIGAZINE, 2026-09-18）
  4. Anthropic

Related