Open-source Python framework for building multi-agent systems with tool use, guardrails, and human-in-the-loop control.
Engineering BlogRenamed Claude Code SDK to Claude Agent SDK to reflect broader vision beyond coding. Claude Code had become used for deep research, video creation, and note-taking internally. Provided framework for building multi-agent systems.
The foundational framework providing orchestration, state management, and execution control for autonomous agents. The SDK abstracts away common agent concerns (retry logic, error handling, tool invocation) so developers can focus on agent behavior and workflows rather than infrastructure, similar to how web frameworks abstract HTTP handling.
An architectural pattern where a central orchestrator agent coordinates multiple specialized worker agents, each optimized for specific tasks. Rather than one powerful agent doing everything, multiple focused agents can handle complex workflows more reliably. The orchestrator decides when to invoke workers, combines their outputs, and handles coordination.
A standardized approach for connecting Claude to external capabilities (APIs, code execution, file systems, MCP connections). The framework handles tool invocation, parameter validation, error handling, and response parsing, enabling agents to reliably use tools without each agent reimplementing integration logic.
Built-in safety mechanisms constraining what agents can do: rate limits, action validators, sandboxed execution environments, and audit logging. These boundaries prevent agents from taking unintended actions while preserving the flexibility needed for complex tasks, balancing autonomy with human oversight.
The infrastructure for running agents reliably at scale: observability and monitoring, error recovery, state persistence, and integration with production systems. Agents in production differ from research agents in requiring high reliability, predictable behavior, and accountability for actions taken.