OpenTelemetry Export and AWS Scheduled Deployments in Claude
Anthropic launched OpenTelemetry export beta for Claude Messages and Managed Agents, plus scheduled agent deployments on AWS. Learn how to configure both.
- OpenTelemetry export is available in beta for Messages API calls and Claude Managed Agents across SDKs and Console destinations.
- Traces can be configured in Metadata only or Full content mode, with default per-span payload truncation at 0.75 MB.
- Scheduled deployments for Claude Managed Agents are now supported on Claude Platform on AWS using IAM role execution identities.
- API spans export directly over HTTPS OTLP/HTTP to endpoints accepting protobuf payloads on port 443.
Two Beta Telemetry and Scheduling Updates for Claude
On 2026-10-14, Anthropic published two major beta additions to the Claude Platform: OpenTelemetry export and Claude Managed Agents scheduled deployments on Claude Platform on AWS. OpenTelemetry export allows engineering teams to pipe structured execution traces from Claude Messages API calls and Managed Agents sessions straight into external observability platforms. At the same time, scheduled deployments expand beyond standard Claude Platform accounts to Claude Platform on AWS, letting organizations automate autonomous recurring agent runs using existing AWS identity boundaries.
These additions address two persistent challenges when managing autonomous agents and LLM calls in production: deep distributed visibility and reliable task cadence. By adopting the vendor-neutral OpenTelemetry standard and aligning with the GenAI semantic conventions, Anthropic enables standard span delivery across modern backends without vendor lock-in. Developers tracking usage costs or token counts across multi-model setups can reference our llm token counts differ between models guide and per run cost cap guide.
How OpenTelemetry Export Works: SDK vs. API Tracing
OpenTelemetry tracing in Claude operates at two distinct layers: SDK tracing and API tracing. You can deploy either approach alone or unite them in a single trace to inspect client-side execution alongside Anthropic backend execution.
| Capability | SDK Tracing | API Tracing | Both Together |
|---|---|---|---|
| Messages API Requests | Yes | Yes | Yes |
| Sessions API Requests | Yes | Yes | Yes |
| Token Usage Counts | Yes | Yes | Yes |
| Individual Model Runs Per Request | No | Yes | Yes |
| Managed Agents Turns | No | Yes | Yes |
| Server Tools (Run by API) | No | Yes | Yes |
| Client Tools (Run by Tool Runner) | Yes | No | Yes |
| Tools in Self-Hosted Sandbox | Yes | No | Yes |
| Auxiliary Calls (e.g. File Uploads) | Yes | No | Yes |
| Configuration Mechanism | Application Code | Claude Console / Admin API | Both |
| Required Access Role | None | Admin | Admin |
| Client Coverage | SDK Clients Only | Any HTTP Client | Any HTTP Client |
| Supported Cloud Platforms | Any | Claude Platform / AWS | Claude Platform / AWS |
The Claude SDK provides trace emission across seven programming languages: Python, TypeScript, C#, Go, Java, PHP, and Ruby. When tracing is enabled inside client code, the SDK emits spans for messages, session requests, and local tool executions handled through the tool runner. However, internal Anthropic steps—such as multi-step server tool executions, individual model calls during agentic loops, and agent turns—are only visible when API tracing is enabled.
Configuring API Telemetry Destinations and Content Modes
API tracing requires setting up an OTLP/HTTP destination. This destination instructs the Claude API where to forward telemetry. Anthropic natively supports several telemetry partner backends, including Arize AI, Braintrust, Databricks, Datadog, Grafana Cloud, Honeycomb, Langfuse, LangSmith, New Relic, PostHog, Sentry, and Splunk. Custom collectors can also be configured if they accept protobuf encoding at an HTTPS URL on port 443.
Each destination and SDK tracer is configured in one of two content modes: Metadata only or Full content.
- Metadata only: The default mode. Spans include operational metadata such as timestamps, status codes, identifiers, token consumption totals, agent descriptions, tool definitions, and Model Context Protocol (MCP) server identifiers. Prompt inputs and completions are omitted.
- Full content: Spans include all raw user prompts, model responses, tool input arguments, and tool outputs, in addition to standard metadata. Under zero data retention policies, the Claude Messages API holds exported payload contents strictly in memory until delivered to your OTLP destination.
- Compliance Restrictions: Organizations governed by HIPAA, FERPA, or K-12 data-handling restrictions automatically fall back to metadata export only for Console destinations, ignoring the Full content setting. SDK-level spans are not constrained by this platform-level override.
Full content spans enforce a default content payload limit of 0.75 MB, which accommodates approximately 200k tokens of standard English text and fits common proxy limits. When payloads exceed this threshold, the API or SDK truncates the content, sets the attribute anthropic.content.truncated to true, and drops the system prompt and earliest messages first. Tool spans drop blocks from the end of their results, while image and document blocks are stripped down to their type metadata. Teams handling up to 1M tokens in context may adjust this ceiling upward in Console settings or via SDK configurations.
curl --fail-with-body -sS "https://api.anthropic.com/v1/organizations/telemetry_destinations" \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "anthropic-beta: telemetry-destinations-2026-08-11" \
-H "content-type: application/json" \
-d '{
"endpoint_url": "https://gateway.example.com/otlp",
"headers": {"authorization": "Bearer YOUR_OBSERVABILITY_API_KEY"},
"traces": {"content_mode": "metadata_only"},
"workspaces": {"type": "all"}
}'
Scheduled Deployments on Claude Platform on AWS
Claude Managed Agents scheduled deployments can now run on Claude Platform on AWS. Scheduled deployments trigger autonomous agent sessions on a predetermined recurring schedule, removing the need for external cron runners or Lambda trigger scripts. Organizations can maintain up to 1,000 scheduled deployments per organization.
On Claude Platform on AWS, every execution must be tied to an IAM role. Before deploying, the target agent must have an execution_identity specifying an AWS role ARN. Creating a scheduled deployment without this attribute configured yields a 400 error.
- Trust Policy Requirements: The role trust policy must permit the aws-external-anthropic.amazonaws.com service principal to execute both sts:AssumeRole and sts:SetSourceIdentity actions.
- Role Permission Requirements: The assumed role must possess permissions allowing aws-external-anthropic:CreateSession, sts:GetWebIdentityToken, and sts:TagGetWebIdentityToken.
- Caller Permissions: The IAM identity configuring the agent must hold iam:PassRole permissions targeting that specific execution role. Note that AWS-managed policies do not grant iam:PassRole automatically.
- Account Boundaries: The execution role must reside within the exact same AWS account associated with the deployment.
Schedules use POSIX cron format with an explicit IANA timezone string. While run timestamps calculate directly from this specification, the platform applies an automated execution jitter between 5 seconds and 9 minutes (calculated as up to 15% of the run interval) to distribute operational load.
curl --fail-with-body -sS "https://api.anthropic.com/v1/deployments?beta=true" \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "anthropic-beta: managed-agents-2026-04-01" \
-H "content-type: application/json" \
-d '{
"name": "Weekly compliance scan",
"agent": "agent_011CYm1BLqPXpQRk5khsSXrs",
"environment_id": "env_01595EKxaaTTGwwY3kyXdtbs",
"initial_events": [
{"type": "user.message", "content": [{"type": "text", "text": "Run the weekly compliance scan."}]}
],
"schedule": {
"type": "cron",
"expression": "0 20 * * 5",
"timezone": "America/New_York"
}
}'
Per-Run Budget Controls and Execution Gotchas
Deployments can define an optional budget ceiling applied independently to every session execution. Passing max_list_cost sets an isolated spending limit per run rather than an aggregate cap over the life of the deployment. For instance, setting an amount value of “2000” USD restricts that specific session run to approximately $20. When that cost threshold is hit, the session transitions into a budget_reached pause state. When planning agent budgets, developers can calculate operational model estimates using our llm api cost calculator.
Several behavioral edge cases must be kept in mind when configuring these systems:
- Daylight Saving Transitions: Wall-clock times operate on literal matching. A scheduled run set during a spring-forward hour that does not exist will skip execution entirely. Conversely, scheduled runs falling into an ambiguous fall-back window trigger twice. Anthropic advises scheduling critical recurring jobs outside 1:00 AM to 3:00 AM local time or fixing the schedule to UTC.
- W3C Distributed Tracing: The Claude API extracts incoming trace context via standard W3C traceparent headers. When traceparent specifies an unsampled flag, the API omits the span. However, Claude Platform on AWS does not yet parse incoming traceparent headers; calls routed through AWS start independent root traces.
- Custom Collector Ports: An OpenTelemetry Collector defaults to listening on port 4318 for OTLP/HTTP traffic. Because Claude API telemetry export strictly targets HTTPS on port 443, self-hosted collectors must use an ingress proxy or load balancer configured for port 443.
- Managed Agents and Zero Data Retention: Messages API calls qualify for zero data retention, meaning content buffers exist solely in memory during trace dispatch. In contrast, Claude Managed Agents is not eligible for zero data retention, and telemetry export does not modify its baseline platform data retention rules.
Implementation and Verification Checklist
Follow this practical sequence to enable and verify telemetry or AWS agent scheduling in your environment:
- Verify Admin Access: Confirm that your Claude Console profile has the Admin role required to create or alter organization telemetry destinations.
- Validate OTLP Endpoint: Verify that your telemetry ingestion proxy accepts HTTPS on port 443 with protobuf serialization before saving the destination URL.
- Select Workspace Scoping: Define destination coverage across All, Selected, or All except workspace configurations.
- Trigger Verification Test: Execute the Test action in the Console destination menu to confirm receipt of the anthropic.test span and check for a Healthy status.
- Inspect Backing IAM Role (AWS): For scheduled deployments on AWS, confirm the target agent holds an execution_identity role ARN with valid trust and session permissions.
- Track Filter Keys: Locate outbound traces in your logging or observability platform by filtering on service.name = claude-platform.