Claude SDK & Cowork — Platform Overview

Claude Agent SDK and Claude Cowork are two expressions of the same underlying agent runtime. The SDK gives developers programmatic control over the loop; Cowork gives non-developers a delegation interface over the same runtime.

Relationship

Claude Agent Runtime (agent loop, tools, hooks, sessions)
         ↙                           ↘
Claude Agent SDK                 Claude Cowork
(custom applications,            (desktop delegation,
 code, API, MCP tools)           files, connectors, skills)

Both share the same fundamental design concerns: bounded tool use, session management, context budget control, and human review at decision points.

SDK: for builders

The SDK is for embedding the agent loop into custom applications. The maturity ladder for SDK learning:

  1. query() — understand the message stream before adding tools
  2. Tool permissions and PreToolUse hooks — control what the agent can do
  3. Streaming UI — turn loop messages into product interfaces
  4. AskUserQuestion / human input — build approval workflows
  5. Custom MCP tools — connect domain APIs
  6. Subagents — decompose complex workflows into specialist roles

Key rule: read-only first, approved writes second, autonomous writes last.

Cowork: for delegators

Cowork is for users who want to hand off complete work packages. The effective pattern:

  • Name the deliverable, inputs, scope, and definition of done
  • Use the smallest folder that contains the work
  • Use plan mode for complex or risky work
  • Review final artifacts like a colleague’s draft
  • Convert recurring work into skills and projects

Key rule: manage Claude like a capable colleague, not a chat partner. Scope is control.

Shared failure modes

FailureSDK manifestationCowork manifestation
Unbounded toolsBroad shell access, data mutationOversized connected folders
Context rotLong session state driftStale history from prior tasks
Missing validationNo result check against criteriaNo human review of final artifact
Vague scopeUnclear deliverable in the promptRe-prompting and rework

Key insight across both

Context is an operating budget — for cost and for reasoning quality. Token management, prompt caching, context compaction, and session lifecycle are not DevOps concerns; they are product design decisions that determine whether an agent is trustworthy in production.