Claude Agent SDK
Claude Agent SDK (formerly Claude Code SDK) is Anthropic’s official SDK to “build production AI agents with Claude Code as a library” — “the same tools, agent loop, and context management that power Claude Code, programmable in Python and TypeScript.” It is the clearest artifact of the model-provider → agent-platform shift this wiki tracks: the harness itself, shipped as a building block. Source: official docs.
What it provides
The Claude Code harness, unbundled: built-in tools (Read/Write/Edit/Bash/Glob/Grep/WebSearch/WebFetch),
hooks (PreToolUse/PostToolUse/SessionStart… — the agent-guardrails insertion points),
subagents (agent-orchestration as a first-class option), MCP (model-context-protocol)
integration, sessions (resume/fork — durable-agents), and permissions. It even loads
.claude/ config — Skills, slash commands, CLAUDE.md memory, plugins — so this
wiki’s threads (skills, orchestration, durability, guardrails) are literally SDK features.
Where it sits
Anthropic now offers a three-rung ladder: the Client SDK (you implement the tool loop) → the Agent SDK (Claude runs the loop in your process) → Managed Agents (Anthropic runs the loop + sandbox). The Agent SDK is the harness-as-library middle rung — strong evidence for the spoke’s “the value moved to everything wrapped around the model, and the model vendor is now selling that wrapper too” thesis.
Billing: a paused metering change
The harness-as-library has an unresolved cost-structure problem. A change announced May 13 would
have, from June 15, metered Agent SDK usage (third-party apps + the claude -p headless command)
separately at Anthropic’s API rates — subscribers getting only a flat monthly credit equal to their
plan price — replacing today’s far more generous weekly subscription caps. On June 15/16 Anthropic
paused it (“for now, nothing has changed”), saying it is reworking the plan. The reprieve looks
temporary: boris-cherny (Head of Claude Code) had said subscriptions “weren’t built for the usage
patterns of these third-party tools.” Full story: anthropic-agent-sdk-billing-pause.
How Anthropic teaches it
agent-sdk-workshop (2026-08, Apache 2.0) is the vendor’s own hands-on course for this SDK, and it
adds something the docs don’t: an order. Four boolean switches in a config file take one agent
through prompt → tools → subagents → memory, so the layers this wiki assembled from separate artifacts
arrive stated as a dependency chain, capability first and persistence last. Its framing — “no code
writing required… you flip switches” — is the harness-as-library claim at full strength: the SDK owns
the loop, the builder supplies configuration. Six breakout scenarios (chief-of-staff, support,
sre-agent, account intelligence) also show where Anthropic expects SDK agents to land, and four of
the six are operations roles rather than coding.
Related
agent-sdk-workshop · agentic-coding-harness · agent-orchestration · durable-agents · agent-guardrails · agent-skills · model-context-protocol · claude-managed-agents · anthropic · adk · agents-md · anthropic-agent-sdk-billing-pause · boris-cherny