Claude Agent SDK+52 Strong positive

GILD Review

GILD's call and the buying facts behind it.

From GILD research

Should we adopt Claude Agent SDK?

Should your team build its agents on the Claude Agent SDK or on its own loop?

30-Second TakePick the Claude Agent SDK if your agents work in files and a shell under approval, and you are committed to Claude.
Best fitTeams committed to Claude that need agents that edit files and run commands under approval, deployed as internal tools or per-task containers.GILD Review
Consider another workflowFor several model providers or an explicit graph, evaluate LangGraph or Pydantic AI. For a loop on OpenAI models, try the OpenAI Agents SDK, and for chat inside a TypeScript web app, the Vercel AI SDK. For a multi-tenant backend, write your own loop on the Claude API or use Claude Managed Agents.GILD Review

Drawn from 117 sources and 6 operator insights

GILD analysis

Claude Agent SDK Review

Thomson Reuters, Spotify, Crystallize and PostHog run production agents on the Claude Agent SDK, which gives a team Claude Code's agent loop, tools, subagents, hooks and permission rules. Each session runs the bundled Claude Code binary as a subprocess with transcripts on local disk, so hosting and session storage take real engineering. It runs Claude models only, and builders measured SDK runs draining a Claude plan's 5-hour window faster than Claude Code in October 2026.

Agent Frameworks
Agentic Development

The decision

Best fit: Teams committed to Claude that need agents that edit files and run commands under approval, deployed as internal tools or per-task containers.
Consider another workflow: For several model providers or an explicit graph, evaluate LangGraph or Pydantic AI. For a loop on OpenAI models, try the OpenAI Agents SDK, and for chat inside a TypeScript web app, the Vercel AI SDK. For a multi-tenant backend, write your own loop on the Claude API or use Claude Managed Agents.
Adoption condition: Pinned SDK versions, a SessionStore for transcripts, turn and budget limits, and permission rules settled before the agent reaches users. On a Claude plan, meter SDK runs against the 5-hour window and treat a quota message as a failed run.

Basis: Anthropic's Agent SDK documentation, help center, and pricing checked Oct 25, 2026, plus 98 public accounts from builders and teams, among them Thomson Reuters, Spotify, Crystallize and PostHog. We ran no performance benchmark.

Buying & deployment factsWhat a team building on it should verify before production traffic grows.

How you buy it

Free SDK

Tokens bill on the Claude API, or against US$100 or US$200 in monthly credits on Max plans.

Source 8, pypi.org, claude-agent-sdk on PyPI: claude-agent-sdk on PyPI Source 1, Anthropic, Agent SDK overview: Agent SDK overview Source 3, Anthropic, Monthly API credits for Max and Team plans: Monthly API credits for Max and Team plans Source 4, Anthropic, Claude API pricing: Claude API pricing

Details

The Python package is free under the MIT license, and the TypeScript package is free under Anthropic's Commercial Terms. Model use is billed per token on the Claude API, or drawn from monthly API credits of US$100 on Max 5x and US$200 on Max 20x. Team plans add US$20 or US$100 per seat, pooled up to US$500 a month.

Cost drivers

US$2 in, US$10 out

Per million tokens on Claude Sonnet 5.5. Cache misses across turns raise the bill.

Source 4, Anthropic, Claude API pricing: Claude API pricing Source 7, Anthropic, Track cost and usage: Track cost and usage

Details

Tokens across every turn and subagent. Claude Sonnet 5.5 costs US$2 per million input tokens and US$10 per million output tokens, and Opus 5.5 costs US$4 and US$20. Cache reads cost a twentieth of the input price, so cache misses across turns raise the bill. The SDK's total_cost_usd field is a client-side estimate.

Where it runs

A process per session

A Python 3.10+ or Node.js 18.14+ service starts a Claude Code binary per session, from 1 GiB RAM.

Source 2, Anthropic, Hosting the Agent SDK: Hosting the Agent SDK Source 5, Anthropic, Agent SDK quickstart: Agent SDK quickstart

Details

In your own process on Python 3.10+ or Node.js 18.14+, which starts a bundled Claude Code binary per session. Anthropic suggests 1 GiB RAM, 5 GiB disk, and 1 CPU per agent as a floor. Models come from the Anthropic API, Amazon Bedrock, Claude Platform on AWS, Google Cloud's Agent Platform, or Microsoft Foundry.

What to verify

139 releases a year

Pin versions, and add a SessionStore so transcripts leave local disk.

Source 6, Anthropic, Persist sessions to external storage: Persist sessions to external storage Source 1, Anthropic, Agent SDK overview: Agent SDK overview Source 8, pypi.org, claude-agent-sdk on PyPI: claude-agent-sdk on PyPI Source 3, Anthropic, Monthly API credits for Max and Team plans: Monthly API credits for Max and Team plans

Details

Transcripts stay on local disk unless a SessionStore is configured. Third-party products cannot offer claude.ai login without Anthropic's approval. The Python package shipped 139 releases in the year to Oct 28, 2026, so pin versions. Plan credits do not cover Claude on Bedrock, Vertex AI, or Foundry.