Claude Agent SDK (www.anthropic.com)

Anthropic has rebranded Claude Code SDK as the Claude Agent SDK to reflect its evolution from a coding assistant into a general-purpose agent platform. The core idea is simple but powerful: give Claude programmatic access to a computer (filesystem, bash, scripts, and tools) so it can gather context, take actions, verify results, and iterate—enabling agents for finance, personal assistants, customer support, deep research, and more. The SDK emphasizes agent loops, tool-first design, code generation for precise outputs (e.g., creating Excel/PPT/Word via generated Python), and built-in features like compaction to summarize long-running context. Technically, the SDK supports two search modes—agentic search (shell-based grep/tail for transparent, accurate retrieval) and semantic search (vectorized chunking for speed but with maintenance tradeoffs)—and first-class subagents for parallel work and isolated contexts. Tools are promoted in the model context to steer action; bash and script execution provide flexible primitives; MCP (Model Context Protocol) offers standardized, OAuth-handled integrations (Slack, GitHub, Google Drive, etc.). Verification strategies include rule-based checks and linting, visual feedback (screenshots/playwright), and optionally LLM-based judges. For the AI/ML community this means a mature set of primitives and best practices for building reliable, auditable, and extensible autonomous agents with real-world integrations and iterative self-correction.