deanlu.ai

ClaudeSDKEcosystem

2 min read

Claude Agent Ecosystem Overview

Anthropic’s Claude agent ecosystem has two entry surfaces — Cowork for delegation-first human users, and the Agent SDK for developers building custom agent applications — unified by the same underlying agent loop.

Ecosystem map

[Human user] → Cowork (projects, skills, connectors, browser, scheduled)
                     ↓
              Shared Agent Loop (model + tools + permissions + sessions)
                     ↑
[Developer] → Agent SDK (query(), hooks, custom MCP tools, subagents)

Cowork: for delegation

Cowork reframes the interaction from chat to delegation. Users hand Claude a complete work package (deliverable, inputs, scope, nuance, review boundary) and Claude plans and executes across files, connectors, and tools. The key discipline is not asking better questions — it’s scoping better work packages.

See ClaudeCowork for components, and ContextBudget for the cost and quality management discipline Cowork demands.

Agent SDK: for developers

The SDK embeds the same loop into custom applications. Developers control which tools are available, how the loop is instrumented, and how results are validated. The demo ladder (hello-world → streaming → approvals → MCP tools → subagents) teaches one production concern at a time.

See ClaudeAgentSDK for components and MCP for the tool protocol.

Shared governance principles

Whether using Cowork or the SDK, the same governance framework applies:

  • Minimal tools per workflow — grant only what the task requires
  • Hooks for audit and policy — AgenticGovernance is architectural, not prompt-level
  • Human review for consequential decisions — Claude executes, humans decide
  • Context as an operating budget — ContextBudget applies to both surfaces

Repository learning order

For SDK developers, the safe learning path:

  1. claude-agent-sdk-python or claude-agent-sdk-typescript — understand the runtime contract
  2. claude-agent-sdk-demos — learn one production concern per demo
  3. claude-quickstarts — understand application packaging
  4. anthropics/skills — patterns for capability reuse (inspect carefully before adopting)

Official repositories reveal the real control surface. Community assets should be treated as design inspiration, not trusted defaults.

Where Claude SDK meets NVIDIA

Both the Claude Agent SDK and NVIDIA’s NeMoAgentToolkit support MCP, making it possible to build agent pipelines that combine Claude’s reasoning with NVIDIA-hosted models or tools. The governance principles are the same across both: separate inference from workflow from runtime policy.

Graph View

Backlinks