Claude Code 2.1.292 adds an effort parameter to the Agent tool, letting you run subagents at five levels β low through max β to balance cost and quality per task. The release also adds a --marketplace flag to plugin install, letting you register the marketplace and install a plugin in a single command.
Claude Code 2.1.290 introduces the "You Should Know" built-in mod β a side agent that catches what both you and Claude might miss. A new managed agents onboarding command lets you bootstrap multi-agent architectures with a single line. VS Code users get a visual Agent Map showing every sub-agent at a glance.
Anthropic added webhooks to Claude Managed Agents. Long-running AI agents now proactively notify your app on key state changes β idle waiting for approval, outcome evaluation complete, errors. With HMAC signing, replay protection, and SDK helpers in 7 languages, this is the practical guide for developers running AI agents in production.
On May 6, 2026, Anthropic added three major capabilities to Claude Managed Agents: Dreaming (self-improving agents that review past sessions), Outcomes (rubric-based grading by a separate evaluator model), and Multi-Agent Orchestration (a lead agent that delegates tasks to specialist sub-agents). AI is evolving from assistant to self-improving teammate.
April 2026 reshaped the AI landscape again. Anthropic opened its Claude Managed Agents public beta and Agent SDK, giving anyone the infrastructure to build autonomous agents. Google supercharged Gemini 2.5 Flash with native audio output and computer use capabilities, while NotebookLM completed its deep integration with Google Classroom β setting a new standard for AI in education.