2026-10-14 Changes confirmed medium confidence

Anthropic Adds Tracing and Scheduled Runs to Managed Agents

The October 14 platform update adds beta OpenTelemetry export and recurring agent deployments on AWS, with optional content capture and explicit timing limits.

Edited by Tyronne Panaino

Anthropic added two operational controls to its Claude platform on October 14: OpenTelemetry export in beta and scheduled deployments for Claude Managed Agents on Claude Platform on AWS. The official release notes position the changes around tracing API and agent activity, then running managed agents on a recurring schedule rather than starting every session manually.

Together, the features move Anthropic's managed-agent service closer to ordinary production operations. Teams can connect model and tool activity to an observability system, while a deployment API can start sessions from a recurring cron schedule. The controls are useful for diagnosis and automation, but Anthropic's own documentation also defines important gaps: exported spans are best effort, content capture is optional, and scheduled execution can move away from the nominal time.

What the new traces cover

Anthropic's OpenTelemetry documentation says the Claude API and SDK can trace Messages API requests, tool uses and Managed Agents sessions. The release notes list SDK span support for Python, TypeScript, C#, Go, Java, PHP and Ruby. API-side tracing is configured through a destination in the Claude Console.

The documented use cases include debugging slow or failed requests, examining token spend and latency, and building evaluations from production traffic. SDK and API tracing do not expose exactly the same view. SDK spans cover activity seen by the application, including client-run tools, while API spans can describe individual model runs, managed-agent turns and server-side tools. Teams can enable both to connect application and service activity, subject to the platform-specific boundaries in the documentation.

Anthropic uses metadata-only capture as the default content mode. In that mode, spans can include identifiers, timing, status, token counts and selected names or descriptions without automatically carrying model prompts and responses. Full-content mode adds prompts and responses. That option changes the data-governance stakes because the receiving observability platform stores exported content under its own access and retention settings.

The traces should not be treated as an infallible ledger. Anthropic says span delivery is best effort and may be incomplete. Sampling can also omit a request's spans. The company points teams that need API-key usage accounting toward its Usage API rather than relying on the trace stream alone. The practical boundary is clear: OpenTelemetry can support investigation and evaluation, but the fetched documentation does not establish a complete audit record.

How scheduled deployments work

Anthropic's scheduled-deployment guide describes recurring Managed Agents sessions controlled through the Deployments API. A deployment combines an agent and environment configuration with an initial event and a schedule containing a cron expression and an IANA time zone. The finest supported schedule is one minute.

For Claude Platform on AWS, the documentation adds an AWS execution-identity requirement so each run creates its session through an IAM role in the customer's account. That makes the scheduled path part of the customer's cloud authorization design rather than a timer detached from deployment permissions.

Nominal schedule time is not an exact execution guarantee. Anthropic says actual runs receive jitter calculated from the interval between runs. The service also supports no more than 1,000 scheduled deployments per organization unless support raises the limit. These constraints matter for workloads that depend on a precise instant or need very large numbers of independent schedules.

A deployment may also set a session budget. Anthropic says the cap is copied onto every run separately rather than accumulated across all runs. That can bound each scheduled session, but it does not by itself prove that the task completed, stayed within an external system's limits or produced a valid result. Run history and application-level outcome checks remain necessary.

Evidence quality and operational limits

All three fetched sources are first-party Anthropic documentation. They establish release timing, supported interfaces, capture modes, scheduling fields and stated service limits. This run found no independent production test of export completeness, trace-delivery reliability, scheduling accuracy, alerting quality or business outcomes.

Internal confidence is therefore medium. The supported conclusion is that Anthropic has shipped beta trace export and recurring Managed Agents deployments with documented controls. Whether those controls satisfy a particular audit, privacy, reliability or automation requirement still depends on configuration and measurement in the user's environment.

Status

Confirmed. Anthropic documents the October 14 availability and mechanics; independent production behavior and operational outcomes remain unverified.

Sources

• Claude Platform release notes
• Anthropic — OpenTelemetry for the Claude API
• Anthropic — Scheduled deployments

Update note: Last reviewed 2026-10-15. We will revise this post if Anthropic changes trace content defaults, delivery guarantees, scheduling availability or deployment limits.

Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.

More Changes coverage