Anthropic

In that hypothetical incident-summary service, the responder sends a ticket, reads the first answer and adds a new diagnostic result. The follow-up needs more than the last sentence. It needs the earlier facts that still apply, along with the new result that may change them.

With ordinary Messages API requests, the application supplies that conversation history each time. These requests are stateless: an earlier call is not a memory the next call can simply consult. The application chooses the context and coordinates the repeated requests and tool results, often called an agent loop. The documentation checked November 3 also describes optional message threads in beta, which can keep the conversation for later requests. That is a separate feature, not a reason to assume an ordinary request remembers the ticket.

If the investigation needs to keep working while the responder attends to something else, Claude Managed Agents supplies a configurable harness and execution infrastructure. The team defines an agent with the model and tools it can use, chooses an environment for those tools to run in, then starts a session for this investigation. The application exchanges messages, tool results and status updates as events; the service persists that event history server-side. Environments can use managed cloud sandboxes or self-hosted sandboxes. This beta service supports ongoing, asynchronous work, with state and execution handled differently from an ordinary Messages request.

Where the application connects matters, too. The API overview distinguishes direct access from cloud routes that integrate billing and identity and access management, or IAM. As documented on November 3, Claude Platform on AWS, Microsoft Foundry and Claude Models on Google Cloud (Anthropic managed) are Anthropic-operated. Amazon Bedrock and Claude Models on Google Cloud (Google managed) are partner-operated. Features and timing vary. An AWS or Google Cloud label alone cannot tell the team which service will handle its investigation.

Sometimes the problem is not the investigation but the client carrying it. An engineer can use Claude Code to read the repository, edit files and run project commands. It is a coding product available in terminal, IDE, desktop and web surfaces, with different account and provider options. The client-repair guide below follows a small example: a failed request being mistaken for useful ticket data.

Before the responder sends that incident note, look at the account receiving it. A personal chat and a commercial integration may use the same model while placing the note under different rules. Paying for a personal plan does not make it a commercial account.

Anthropic's consumer policy covers Free, Pro and Max, including Claude Code through those accounts. Chats and coding sessions may be used for model improvement when the user allows it, supplies feedback or otherwise opts in. Conversations flagged for safety review may also be used for safeguards work. The note's subject has not changed; the account and the choices made there change how it may be used.

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 provided through feedback or other training opt-ins may be used. That answers the training question. It does not, by itself, say when a saved note disappears.

The API retention reference checked November 3 says conversation content is not retained by default, except for listed models requiring 30-day retention. Zero data retention, or ZDR, has model- and feature-specific eligibility and must be enabled separately for each organization. A stateful Managed Agents session keeps information needed to continue the investigation and is not ZDR-eligible, even with a self-hosted sandbox. Files, jobs and execution state have their own storage rules; flagged content and legal holds can also change retention. On Amazon Bedrock and Google-managed Claude Models on Google Cloud, the cloud provider is the data processor, so review that provider's controls. The Anthropic-managed Google route belongs to the separate Anthropic-operated scope.

There is another policy wording to keep beside that feature reference. Anthropic's commercial retention policy describes API deletion within 30 days, with exceptions for stateful services, agreed arrangements, policy enforcement and legal requirements. Saved commercial conversations remain available so work can continue. Neither statement supplies one storage duration for every product. For this note, identify the actual account, model, feature and route before deciding which promise applies.

The responder starts reading the incident summary as its first sentences appear. Then the text stops halfway through an explanation. The screen has shown progress, but the client still has a question to answer: did generation finish, or did the connection deliver only part of it?

Streaming sends those pieces through server-sent events, or SSE. An HTTP 200 status can arrive before an error later in the stream, so that initial successful response is not enough. The streaming reference describes updates to content blocks, updates to the message as a whole and the final message-stop event, as well as errors within the stream. The client records completion and inspects the resulting message and stop reason. A paragraph on screen is partial output; how the message ended tells the application what it actually received.

Other requests may be waiting before any text appears. Claude's Messages rate limits separately measure requests, input tokens and output tokens per minute for each model class. Tokens are the pieces of text the model processes. Organization limits can sit alongside workspace controls. Requests crowded into a short interval can hit a limit despite an acceptable-looking minute total, and sharp traffic increases can hit acceleration limits. These are usage ceilings, not a promise of capacity.

For throttling, the client follows the retry-after header when it is present. A tier spend-cap 429 is different: it has no retry timing and keeps failing until access resumes. Monthly spending restrictions are a separate control. Sending the same request again cannot restore access that the cap has paused.

The error reference distinguishes rate limits, temporary overload and other failures. Official SDKs, the libraries an application uses to call the service, retry transient failures twice by default, with increasing delays and respect for retry-after. If a surrounding job runner also retries, count the attempts the SDK already makes. Keep the request ID with a failure so support can locate that call. For the waiting responder, the useful result is not an indefinitely busy screen: it is a complete summary or an honest indication that this attempt did not finish.