Anthropic · Claude Code · Claude Cowork · Agent SDK
Claude 智能体、子智能体与动态工作流
仓库级编码智能体、业务用 Cowork，以及脚本化的多智能体编排 — 配上 Workstation 团队能够落地运行的治理
Claude 不再只是对话模型。Claude Code 在终端和 IDE 中运行能承担长任务的编码智能体。Claude Cowork 处理演示文稿、文档和排好期的知识工作。Agent SDK 增加具名子智能体和动态工作流，可以铺开数十个专门的工作者。Workstation 帮助您为每个团队选对使用面，并在进入生产之前放上审批门、MCP OAuth 以及 Ring Promoter 推进。
Subagents
Isolated workers with tool scopes
Workflows
Scripted fan-out & verification
MCP
Tools, data & connectors
Cowork
Business artefacts & routines
Watch: Building more effective AI agents
点击以在新窗口中打开
Building more effective AI agents — Anthropic
What are Claude agents at Workstation?
Claude agents are Anthropic’s productised ways to plan, use tools, and finish multi-step work — not one-shot chat replies. Engineering teams live in Claude Code. Ops and knowledge teams use Claude Cowork. Builders embed the Claude Agent SDK with subagents and dynamic workflows. Workstation maps each surface to owners, risk classes, and promotion paths so demos become durable operating models.
- Claude Code: Terminal and IDE coding agents for repo-scale edits, tests, PRs, and long-running tasks
- Claude Cowork: Desk-to-phone business workflows for decks, docs, spreadsheets, and scheduled work
- Subagents: Named workers with isolated context and restricted tools for focused subtasks
- Dynamic workflows: Claude-written orchestration scripts that fan out, verify, and synthesise at scale
- Governance: MCP OAuth, Vault leases, human approval gates, and Ring Promoter health-gated deploys
Claude agent surfaces
Claude Code
Anthropic’s agentic coding product for engineering and platform teams. It understands repositories, edits across files, runs tests, and can work headlessly for CI bots and Dispatch-style automation.
- Repo-aware multi-file coding agents
- Hooks, MCP servers, and IDE integrations
- Headless / Dispatch automation
- Remote and container workflows
- Pairs with Claude models on Platform and Bedrock
Claude Cowork
Cowork is Claude for knowledge work: connect data, schedule recurring tasks, and deliver artefacts for human review — without hand-authoring agent files for every job.
- Scheduled recurring business tasks
- Decks, documents, and spreadsheets
- Automatic parallel workstreams
- Plugins that bundle skills and connectors
- Cross-device task check-in
Subagents & Agent SDK
Define specialised subagents in code or `.claude/agents/` markdown. Isolate context, restrict tools, and let the main agent delegate — or call them by name.
- Programmatic AgentDefinition / agents map
- Filesystem-based agent definitions
- Parallel analysis with tool allowlists
- Python and TypeScript Agent SDK
- MCP for databases, browsers, and internal APIs
Dynamic workflows
When a few subagents are not enough, Claude writes a JavaScript orchestration script. The Workflow runtime fans out, stores intermediate results in variables, and returns a verified synthesis — ideal for audits, migrations, and cross-checked research.
- Scripted parallel() and pipeline() orchestration
- Bundled /deep-research workflow
- Ultracode effort for automatic workflow planning
- Resumable background runs
- Adversarial verification passes
Where Claude agents fit
Map Claude Code, Cowork, subagents, and workflows to real team jobs
Engineering
Repo-scale coding & PR loops
Claude Code plans changes, edits files, runs tests, and opens PRs. Subagents review security and quality in parallel before humans merge.
Key Benefits:
- Faster feature delivery
- Consistent review depth
- CI-friendly Dispatch bots
5x faster multi-file changes
Fewer missed review angles
Headless PR bots
Platform & SRE
Codebase audits with workflows
Dynamic workflows fan out across packages for dependency, config, and security sweeps — then verify claims before the report lands.
Key Benefits:
- Repeatable audit scripts
- Cross-checked findings
- Scoped tool permissions
Dozens of agents per run
Verifier passes
Allowlisted tools only
Ops & PMO
Cowork for decks and status packs
Claude Cowork connects sources, builds status decks and spreadsheets, and schedules recurring updates for human review.
Key Benefits:
- Scheduled artefacts
- Less copy-paste reporting
- Human approval before send
Hours saved per week
Recurring routines
Human gate retained
Enterprise builders
Custom agents on the SDK
Embed the Claude Agent SDK behind MCP, Vault leases, and Workstation gateways so internal tools stay authenticated and auditable.
Key Benefits:
- MCP connectors
- OAuth and Vault leases
- Ring Promoter promotion
Least-privilege tools
Session + durable stores
Health-gated rings
Claude agent architecture
Surfaces, orchestration, tools, and controls Workstation recommends
Claude Code runtime
Interactive and headless agent loop with repo tools, hooks, and remote execution options
CLI / IDEDispatchHooksRemote
Agent SDK
Programmatic query loop with Agent tool, allowedTools, and custom tool registration
PythonTypeScriptAgent toolMCP
Subagents
Isolated workers with descriptions Claude uses to auto-delegate focused work
Tool scopesParallel runsNamed agents
Dynamic workflows
Background JavaScript orchestration for large fan-out and verification patterns
Workflow tool/deep-researchUltracode
Claude Cowork
Knowledge-work surface with connectors, plugins, and automatic parallelisation
SchedulesArtefactsPlugins
Workstation controls
MCP OAuth, Vault leases, WSL Proxy edge, and Ring Promoter promotion for production
ApprovalsAuditRings
Claude agents vs traditional chat AI
| 方面 | Traditional chat | Claude agents | 提升 |
|---|---|---|---|
| Work style | Single-turn answers | Multi-step tool-using agents | End-to-end delivery |
| Coding | Snippet suggestions | Repo-scale Claude Code agents | Shipable changes |
| 规模 | One context window | Subagents + workflows | Fan-out with verification |
| Business work | Draft in chat | Cowork artefacts & routines | Scheduled outcomes |
| 企业版 | Copy-paste integrations | MCP + governed promotion | Auditable production |
Latest Claude agent updates
2026
Dynamic workflows in Claude Code
Claude writes orchestration scripts that fan out subagents for audits, migrations, and cross-checked research.
Claude Code docs2026
Claude Cowork for business teams
Desk-to-phone workflows for decks, docs, and scheduled knowledge work without hand-written agent files.
Anthropic product2026
Agent SDK subagents
Define specialised workers with tool scopes and descriptions Claude uses to delegate automatically.
Agent SDK docs2026
Workstation teams guide
Map Claude Code, Cowork, and ChatGPT agents to engineering, ops, GTM, and finance with shared controls.
Workstation articlesAdopt Claude agents with Workstation
We help you choose surfaces, wire MCP securely, and promote safely — Anthropic pricing stays with Anthropic
Pilot
Scoped engagement
One team, one Claude surface, clear success metrics
- Claude Code or Cowork pilot
- MCP connector review
- Approval-gate checklist
- 2-week working session
最受欢迎
Team rollout
Programme per quarter
Multi-team playbooks with SDK patterns and Ring Promoter
- Subagent / workflow blueprints
- Vault + OAuth patterns
- Ring Promoter promotion
- Operator training
企业版
Custom contact
Bedrock / Platform routing, private networking, and SLAs
- Private AI options
- WSL Proxy edge
- Compliance packaging
- Dedicated account team
Frequently Asked Questions
Claude chat answers questions. Claude Code is an agentic coding product that reads your repo, edits files, runs commands, and completes multi-step engineering work in the terminal or IDE — including headless Dispatch for automation.
Subagents are separate agent instances the main agent can spawn for focused work. You define them in the Agent SDK or as markdown files under .claude/agents/, give them clear descriptions and tool allowlists, and Claude delegates when the description matches.
Use a few subagents for adaptive delegation inside a conversation. Use dynamic workflows when you need dozens of workers, repeatable orchestration in a script, and verification passes that should not overflow the main context window.
Cowork targets business and knowledge work — decks, docs, spreadsheets, scheduled tasks — with connectors and automatic parallelisation. It is the non-terminal sibling of the same agent architecture as Claude Code.
We map the right Claude surface to each team, secure MCP tools with OAuth and Vault leases, put WSL Proxy in front of agent traffic, and promote changes through Ring Promoter rings so agent demos survive production.
准备好在治理之下运行 Claude 智能体了吗？
Workstation 帮助工程与业务团队采用 Claude Code、Cowork、子智能体和工作流，而不跳过审批门。