Choose a surface your team can operate

Learning objective

Choose the delivery surface — end-user Claude apps, developer tooling, direct API integration, or an agent framework or hosted agent runtime — by users, control, and integration needs.

Core idea

The execution-loop choice is only part of delivery design. You also decide where people work, what triggers the task, where outputs land, and which team operates the runtime. A Claude app, Claude Code, a custom API service, and a hosted agent harness can all involve Claude while exposing different integration and ownership boundaries.

Start with the user population and the required workflow. Then compare control and integration needs. Choose the surface that satisfies those needs with responsibilities your team can actually operate. A convenient demonstration surface need not be the production delivery surface.

How it works

Identify the user and the destination

A person preparing a briefing needs an interactive place to read and revise. A developer modifying a repository needs code, tools, and a review workflow. A backend processing every incoming record needs a callable interface, application identity, and an explicit output destination. These requirements are stronger selection signals than the label “assistant.”

For end-user work, consider Claude apps where the organization already governs user activity and content. Anthropic’s Compliance API documentation distinguishes claude.ai organizations, where users chat and store content, from Claude Console organizations, where users manage API workloads. This is a concrete organizational boundary: procurement of app access does not itself define a production backend integration.

For developer work, Claude Code provides terminal, IDE, desktop, and browser surfaces and can read code, edit files, and run commands. Select the surface that fits the team’s development environment and approval process. Giving developers an interactive coding tool is a different delivery decision from embedding a coding agent in a product service.

Match the integration to its owner

Direct API integration fits an application that needs to construct requests, control the surrounding workflow, and deliver results through its own interface. The client SDK gives direct access to the Claude API. Your team retains the application integration and outcome contract even if it adopts a helper abstraction.

Direct integration can also reach Claude through a cloud platform when its identity, billing, or data-residency requirements shape the design. Check the required surface on that platform’s page. For example, Claude in Amazon Bedrock serves the Messages API at /anthropic/v1/messages on AWS-managed infrastructure, but does not support Claude Managed Agents, the Compliance endpoint, the MCP connector, or structured outputs. Claude Platform on AWS is the Anthropic-operated alternative. Do not infer the same limitations for Google Cloud or Microsoft Foundry.

An agent framework can simplify calls, tool definitions, response parsing, and chaining. Anthropic warns that extra abstraction can obscure prompts and responses and encourage unnecessary complexity. Select one only when its supplied behavior removes work your team would otherwise implement, and confirm that the team can inspect the underlying control path. A framework name does not settle where the process runs.

The Agent SDK embeds Claude Code’s capabilities in a Python or TypeScript application, in a process you operate. It is suitable when you need those capabilities programmatically and want control over the operating environment. Do not repeat the manual API loop around it; your delivery decision concerns deployment and integration responsibilities, assuming the harness distinction.

Separate hosted harness from sandbox location

Claude Managed Agents supplies an Anthropic-hosted harness configured through the Claude API. Its documented concepts are an Agent, an Environment, a Session, and Events. An agent carries model and behavior configuration; an environment determines where execution runs; a session performs work; events exchange user turns, tool results, and status updates.

The TypeScript quickstart creates these resources through client.beta.agents.create(), client.beta.environments.create(), and client.beta.sessions.create(). This interface makes the delivery concrete: your backend initiates and observes a session rather than launching its own Agent SDK process.

Managed Agents supports an Anthropic-managed cloud sandbox or a self-hosted sandbox on your infrastructure. Do not equate a hosted harness with every file and tool executing in Anthropic’s sandbox. Conversely, a self-hosted sandbox does not establish that all model processing stays on your infrastructure. Map both boundaries before approving a data constraint.

Managed Agents stores conversation history, sandbox state, and outputs server-side. It is not eligible for Zero Data Retention (ZDR) or HIPAA Business Associate Agreement (BAA) coverage. Sessions and uploaded files can be deleted separately through the API. The API and data retention page gives per-feature eligibility and applies these exclusions to self-hosted sandboxes too. The organization’s compliance owner determines whether that boundary fits its obligations; choosing a self-hosted sandbox does not remove the exclusion.

Managed Agents is beta. All its endpoints require anthropic-beta: managed-agents-2026-04-01; the SDK sets that header automatically. Its overview also names availability on Claude Platform on AWS with differences in features and session behavior. Evaluate the specific route; that statement is not evidence of equivalent availability on every cloud platform.

Key distinctions and trade-offs

The surfaces divide by intended workflow and operating responsibility:

App governance can be programmatic: Anthropic’s Compliance API offers access to Enterprise activity and content. That does not make app chats interchangeable with your custom service contract. Likewise, hosted execution removes harness work but does not transfer acceptance of business outputs to the vendor.

Architecture / decision scenario

A technology company has six platform engineers and three requests. Account managers need help drafting technical briefings. Developers want assistance fixing repository defects. A customer-facing service must investigate incident evidence asynchronously and return an approved report to the company’s portal. The engineers cannot operate an additional agent-process fleet this quarter.

You select organization-governed Claude apps for the account managers’ interactive work and Claude Code in the developers’ existing environment for repository tasks. Neither choice becomes the backend for customer incidents merely because employees already use it.

For incidents, first compare a fixed direct API workflow with an agent. If supplied evidence is sufficient, keep the direct integration. If the proof of concept demonstrates that variable investigative steps are necessary, compare an operated Agent SDK process with Managed Agents. The staffing constraint favors the hosted harness, conditional on beta acceptance and compliance-owner approval of server-side retention without ZDR or HIPAA BAA coverage. If the company requires the AWS-managed Bedrock route, Managed Agents is unavailable there.

The portal backend creates and observes Managed Agents sessions and validates the resulting report before release. If a requirement mandates tools on company infrastructure, examine a self-hosted sandbox without claiming it solves every processing-location constraint. If that route cannot satisfy the acceptance contract, keep the direct service or defer the agent rather than silently weakening the requirement.

Common exam traps

• Choosing a surface from the demonstration instead of the production trigger and output destination.
• Treating Claude Code as identical to embedding the Agent SDK in a service.
• Treating an agent framework as a hosting service.
• Assuming a hosted session store fits a zero-retention requirement.
• Assuming a self-hosted sandbox establishes the location of all processing.
• Ignoring Managed Agents’ beta header or assuming platform parity.

What to remember

Select by users, control, integration, and operating capacity. Distinguish interactive apps and developer tools from callable services, and distinguish an operated harness from a hosted one. Record sandbox location, platform support, and feature status separately from who accepts the final output.

Knowledge check

References

• Anthropic, Compliance API
• Anthropic, Claude Code overview
• Anthropic, Agent SDK overview
• Anthropic, Building effective agents
• Anthropic, Claude Managed Agents overview
• Anthropic, Managed Agents quickstart
• Anthropic, Claude in Amazon Bedrock
• Anthropic, API and data retention