Anthropic

The Messages API gives an application direct model access. It is stateless: a follow-up request supplies the conversation history it needs. That makes the application responsible for assembling context and coordinating its own agent loop. In an illustrative incident-summary service, this means choosing which ticket messages and diagnostic results enter each request, rather than assuming the API remembers an earlier turn.

Claude Managed Agents supplies a configurable agent harness and execution infrastructure. An agent defines the model and tools; an environment defines the sandbox; a session performs the task and exchanges events with the application. It supports managed cloud sandboxes and self-hosted sandboxes, while persisting event history server-side. This is beta infrastructure for ongoing, asynchronous work, with a different state and execution model from a Messages request.

Deployment route also matters. The API overview distinguishes direct Anthropic access from cloud-provider access with integrated billing and identity and access management (IAM). Claude Platform on AWS and Microsoft Foundry are Anthropic-operated; Amazon Bedrock and Google Cloud are partner-operated. Feature availability and timing vary. An AWS deployment therefore needs the particular service name, not just an AWS label, when comparing capabilities or support.

Claude Code is the coding product an engineer uses to read a repository, edit files and run project commands. It is available through terminal, IDE, desktop and web surfaces, whose account and provider options differ. The Claude Code guide linked on this page develops a concrete client-repair exercise.

Anthropic separates consumer and commercial data policies. Its consumer policy covers Free, Pro and Max, including Claude Code used through those accounts. Chats and coding sessions may be used when the user allows model improvement, submits feedback or otherwise opts in; flagged conversations may also be used for safeguards work. Paying for a personal plan does not place it under the commercial policy.

For commercial offerings, including the API and Claude for Work, Anthropic says it does not train on chats or coding sessions unless the customer participates in its Development Partner Program. Materials explicitly supplied through feedback or other training opt-ins can be used. Training use and the storage needed to provide a feature are separate questions.

The current API retention reference says conversation content is not retained by default, except for Covered Models requiring 30-day retention. It lists model- and feature-specific eligibility for zero data retention (ZDR), which is enabled separately for each organization. Stateful Managed Agents sessions are not ZDR-eligible, including self-hosted sandboxes. Features that store files, jobs or execution state have their own rules. Flagged content and legal holds can also change retention. On Bedrock and Google Cloud, the cloud provider is the data processor, so its controls need separate review.

Anthropic's broader commercial retention policy separately describes API deletion within 30 days, with exceptions for stateful services, agreed arrangements, policy enforcement and legal requirements. Saved commercial conversations remain available for continuation. That wording should be read alongside the current feature reference, rather than converted into one storage duration for every Claude product.

Claude's Messages rate limits measure requests per minute, input tokens per minute and output tokens per minute separately, for each model class. Organization-level limits can coexist with workspace controls. A short burst can trigger a limit even when a minute's total appears acceptable, and a sharp traffic increase can hit acceleration limits. For throttling, use the retry-after header when present. A tier spend-cap 429 has no retry timing and persists until access resumes. These ceilings describe allowed usage, not guaranteed capacity; monthly spending restrictions are a separate control.

The error reference distinguishes rate-limit errors from temporary overload and other failures. Official SDKs retry transient failures with exponential backoff, twice by default, and honor the retry-after header. That behavior matters when adding retries in a surrounding job runner: the application needs to account for attempts the SDK already makes. Keep the request ID with the failure so support can locate the specific call.

Streaming returns output incrementally through server-sent events (SSE). An HTTP 200 can precede a mid-stream error. The streaming reference describes content-block updates, top-level message updates and the final message-stop event, plus errors within the stream. Record whether the stream completed and inspect the resulting message and stop reason. In that illustrative incident-summary interface, text appearing on screen establishes partial output; the final message establishes how generation ended.