Evaluation record · claude-agent-sdk

Claude Agent SDK

v0.3.x

Anthropic

Agentagent-sdkanthropicframework

Strong

About This Agent

SDK exposing Claude Code's production agent harness (tool loop, permission system, subagents, MCP) for building general-purpose agents in TypeScript and Python. Renamed from Claude Code SDK in September 2025.

Last Evaluated: July 9, 2026

Official Website

Trust Vector Analysis

Dimension Breakdown

🚀Performance & Reliability

task completion accuracy

Evaluation of agents built on the SDK harness across coding and non-coding tasks

Evidence

Claude Agent SDK overview — Reuses the same battle-tested agent harness that powers Claude Code, inheriting its task-completion behavior

highVerified: 2026-07-09

tool use reliability

Testing of built-in tool loop, custom tool definitions, and MCP server integration

Evidence

Claude Agent SDK documentation — Built-in file, bash, and web tools plus custom tools and MCP servers with a managed agentic tool loop

highVerified: 2026-07-09

multi step planning

Long-horizon agent task evaluation using the SDK's managed loop and compaction

Evidence

Claude Agent SDK documentation — Automatic context management and compaction enable long-horizon multi-step agent runs

highVerified: 2026-07-09

memory persistence

Review of session resume, memory file support, and context compaction behavior

Evidence

Claude Agent SDK sessions documentation — Session persistence and resume, CLAUDE.md-style memory files, and automatic compaction of long contexts

mediumVerified: 2026-07-09

error recovery

Observed recovery from tool errors and failed commands in SDK-built agents

Evidence

Claude Agent SDK documentation — Harness retries and self-corrects on tool failures; hooks allow custom failure handling

mediumVerified: 2026-07-09

agent collaboration

Testing of programmatic subagent definition and parallel delegation

Evidence

Claude Agent SDK subagents documentation — First-class subagents with isolated contexts, custom prompts, and per-agent tool restrictions

highVerified: 2026-07-09

🛡️Security

tool sandboxing

Review of inherited sandbox configuration and bash isolation options

Evidence

Claude Agent SDK permissions documentation — Inherits Claude Code's sandboxing support, including OS-level filesystem and network isolation for bash

highVerified: 2026-07-09

access control

Assessment of permission rules, programmatic approval callbacks, and hook-based gating

Evidence

Claude Agent SDK permissions documentation — Programmatic permission modes, allow/deny rules, canUseTool callbacks, and hooks for fine-grained control

highVerified: 2026-07-09

prompt injection defense

Review of documented mitigations and developer responsibility for untrusted input handling

Evidence

Claude Agent SDK documentation — Permission gating and sandboxing mitigate injection impact, but defenses against untrusted content depend on developer configuration

mediumVerified: 2026-07-09

data isolation

Architecture review of session/subagent context isolation in self-hosted deployments

Evidence

Claude Agent SDK sessions documentation — Sessions and subagents maintain isolated contexts; deployment isolation is the integrator's responsibility

mediumVerified: 2026-07-09

open source transparency

License and source availability review of npm/PyPI packages

Evidence

@anthropic-ai/claude-agent-sdk on npm — Distributed under Anthropic Commercial Terms (proprietary); SDK wrapper code inspectable but the underlying harness is closed

highVerified: 2026-07-09

🔒Privacy & Compliance

data retention

Review of Anthropic API retention terms and local-state architecture

Evidence

Anthropic privacy policy — API traffic not used for training by default; agent state and transcripts stay on the integrator's infrastructure

mediumVerified: 2026-07-09

gdpr compliance

Compliance certification review including cloud-provider routing options

Evidence

Anthropic Trust Center — SOC 2 Type II and GDPR-aligned DPA available; Bedrock/Vertex routing supports regional data residency

mediumVerified: 2026-07-09

third party data sharing

Data flow analysis across supported model endpoints

Evidence

Claude Agent SDK documentation — Data flows only to the configured Claude endpoint (Anthropic API, Bedrock, or Vertex AI)

mediumVerified: 2026-07-09

local deployment option

Deployment options assessment for self-hosted harness with cloud-only inference

Evidence

Claude Agent SDK documentation — Harness runs entirely on your infrastructure, but inference requires Claude via API, Bedrock, or Vertex; no local models

highVerified: 2026-07-09

👁️Trust & Transparency

documentation quality

Documentation completeness review across both language SDKs

Evidence

Claude Agent SDK documentation — Thorough TypeScript and Python guides covering permissions, subagents, MCP, sessions, and deployment

highVerified: 2026-07-09

execution traceability

Review of message stream observability and hook-based audit trails

Evidence

Claude Agent SDK documentation — Streaming message API surfaces every tool call and result; hooks enable custom audit logging

highVerified: 2026-07-09

decision explainability

Assessment of streamed reasoning visibility and permission-request context

Evidence

Claude Agent SDK documentation — Agent reasoning, tool intents, and permission requests are exposed in the streamed transcript

mediumVerified: 2026-07-09

open source code

Open source assessment of SDK packages and underlying harness

Evidence

@anthropic-ai/claude-agent-sdk on npm — Proprietary license (Anthropic Commercial Terms); package code visible but not open source

highVerified: 2026-07-09

community activity

Community engagement analysis via package downloads, release cadence, and ecosystem projects

Evidence

Claude Agent SDK ecosystem — High npm download volume, frequent releases, and growing ecosystem of SDK-built agents since the 2025-09-29 rename

mediumVerified: 2026-07-09

⚙️Operational Excellence

ease of integration

Integration complexity assessment from install to first working agent

Evidence

Claude Agent SDK quickstart — Single package install for TypeScript or Python; minimal code to launch a fully tool-equipped agent

highVerified: 2026-07-09

scalability

Assessment of horizontal scaling patterns and API rate limit constraints

Evidence

Claude Agent SDK documentation — Stateless harness instances scale horizontally; throughput bounded by Claude API rate limits

mediumVerified: 2026-07-09

cost predictability

Pricing model analysis of token-based costs for long-running agents, including the June 2026 subscription-credit change

Evidence

Anthropic API pricing — Free SDK; costs are pay-as-you-go Claude tokens, which vary widely with agent autonomy and task length

Claude Agent SDK documentation — Since 2026-06-15, Agent SDK and non-interactive (claude -p) usage on subscription plans draws from a separate monthly Agent SDK credit ($20 on Pro, $100 on Max 5x, $200 on Max 20x), capping subscription-based SDK spend

highVerified: 2026-07-09

monitoring capabilities

Review of built-in usage reporting and integration points for external observability

Evidence

Claude Agent SDK documentation — Per-message usage/cost reporting and hooks for custom telemetry; full APM requires external tooling

mediumVerified: 2026-07-09

production readiness

Maturity assessment based on shared production harness and release stability

Evidence

Claude Agent SDK overview — Same harness running Claude Code in production at scale; renamed and generalized for agents on 2025-09-29

highVerified: 2026-07-09

Strengths

+Production-proven harness shared with Claude Code (tool loop, compaction, retries)
+Fine-grained permission system with allow/deny rules, callbacks, hooks, and sandboxing
+First-class subagents and MCP support for composable agent systems
+Both TypeScript and Python SDKs with feature parity
+Automatic context management enables long-running agents
+Runs on Anthropic API, Amazon Bedrock, or Google Vertex AI

Limitations

!Proprietary license (Anthropic Commercial Terms); not open source
!Claude models only; no third-party or local model support
!Token costs for autonomous agents can be hard to forecast
!Heavier runtime footprint than thin LLM client libraries
!Security posture depends on integrator configuring permissions and sandboxing correctly

Metadata

license: Proprietary (Anthropic Commercial Terms)

supported models

0: Claude via Anthropic API

1: Claude via Amazon Bedrock

2: Claude via Google Vertex AI

programming languages

0: TypeScript

1: Python

deployment type: Self-hosted SDK (cloud inference)

tool support

0: Built-in file, bash, and web tools

1: Custom tools

2: MCP servers

3: Hooks

4: Subagents

first release: 2025 (as Claude Code SDK); renamed Claude Agent SDK 2025-09-29

pricing: Free SDK; pay-as-you-go Claude token costs via API, or monthly Agent SDK credits on subscription plans since 2026-06-15 ($20 Pro / $100 Max 5x / $200 Max 20x)

packages

0: @anthropic-ai/claude-agent-sdk (npm)

1: claude-agent-sdk (PyPI)

Use Case Ratings

code generation

Inherits Claude Code's coding harness; ideal for building custom coding agents and CI bots

research assistant

Strong general-agent harness for research workflows with web, file, and MCP tools

customer support

Permission system and subagents suit support automation; requires custom integration work

data analysis

Bash and file tools enable scripted analysis agents with auditable execution