Anthropic now exports OpenTelemetry traces from Claude Managed Agents, the Messages API, and the Claude SDKs to a customer’s own observability tools in public beta. Managed Agents traces omit the model’s reasoning and can drop spans, so the record of why an agent acted must come from the builder.
What Is Covered in This Article:
- What Anthropic’s October 2026 OpenTelemetry trace export covers and the two content modes it offers.
- What the exported traces contain for the Messages API and for Managed Agents, and what they leave out.
- How the release compares with agent tracing from AWS, Google, and OpenAI, and how New Relic, Dynatrace, and Splunk have responded.
- What developers, platform and security teams, buyers, and Anthropic should do next.
The News: On October 14, 2026, Anthropic released OpenTelemetry (OTel) trace export for the Claude Platform in public beta. It covers Claude Managed Agents, the Messages API, and the Claude SDKs. Anthropic says “every agent turn, model call, and tool use arrives in your observability tool as a standard OpenTelemetry trace, with latency, tokens, and IDs.” In OpenTelemetry terms, a trace is the full record of a request through a system, and a span is a single step within it, such as a model request or a tool use. Managed Agents and the Messages API require an Admin to add a destination in the Claude Console. The SDKs export traces once a developer registers a tracer provider in the application. The traces follow OpenTelemetry’s semantic conventions for generative AI.
Anthropic names 12 partners across three uses: debugging agents next to logs and metrics (Grafana Cloud, Honeycomb, Sentry, Splunk), turning production traces into evals (Arize AI, Braintrust, Datadog, Langfuse, LangSmith, New Relic, PostHog), and tracking token usage across an organization (Databricks Unity Gateway). Four customers are quoted: Foxglove, Ramp, Sterling Labs, and FloQast.
Anthropic’s documentation describes two content modes. “Metadata only” is the default and carries identifiers, timing, status, token counts, and text the customer chooses. “Full content” adds prompts and responses. For organizations with a HIPAA, FERPA (student records), or K-12 restriction with Anthropic, destinations export metadata only “regardless of their content mode.” Details are in Anthropic’s announcement.
Claude Agent Observability Can’t Be the Audit Record on Its Own
Analyst Take: Anthropic‘s Claude agent observability provides teams with a standard trace of what their agents did in the observability tools they already use. For Claude Managed Agents, the runtime Anthropic hosts, that trace can’t be the audit record on its own. The record of why an agent acted has to be written by the builder.
Most organizations can’t yet show what their agents did. In The Futurum Group’s 2H 2026 Software Lifecycle Engineering Decision Maker Survey (n=839, fielded June 2026), 45% log what their AI agents do, and 12% evaluate agents before promoting them. Three-quarters experienced a production incident in the past year in which AI was a contributing factor, and 36% of those incidents involved agents taking unintended or out-of-scope actions.
Anthropic pitches tracing as the step before customers widen an agent’s access. On the Messages API, the traces can support that. On Managed Agents, they show what the agent did but leave out why, a gap we take up from the buyer’s side below. Anthropic’s sequence matches our view that agent deployment moves at the pace of what organizations can observe, control, and prove, and Managed Agents traces cover only the first of those.
Claude Agent Observability Falls Short of an Audit Record for Managed Agents
The Messages API path meets a demanding standard. Anthropic’s launch post quotes Ankush Sharma, Head of Infrastructure at FloQast: “Agents that touch a company’s books have to show their work, the same way an accountant does.” His team sees “the prompt version, tool calls, latency, and token spend,” so “when an auditor asks what an agent did and why, we can answer from one timeline.”
On that path, Claude agent observability can record the model’s thinking. With Full content on, Anthropic’s span reference maps the model’s thinking to a reasoning field.
On Managed Agents, a turn carries the agent’s input, messages, tool calls, and tool results, but the thinking is “Not in Managed Agents spans, because the agent.thinking event doesn’t include the thinking content.” The runtime doesn’t expose the thinking, so the export has nothing to send. An auditor gets the actions and outputs without the reasoning that connects them.
For HIPAA, FERPA, or K-12 customers, spans sent from Anthropic’s servers carry only metadata. Spans from the customer’s own SDK aren’t affected, but a Managed Agents turn runs on Anthropic’s servers, so its spans show which tools ran, in what order, how long they took, and whether they succeeded, but no prompts, messages, or results.
Delivery adds a third limit. Anthropic documents API export as “Best effort: A span can be lost before delivery.” The API can’t resend a lost span, so the export can support an investigation but can’t serve as evidence by itself.
The Why Has to Be Built In, Not Exported
Thinking text still helps when debugging, but it can’t prove a decision. In April 2025, Anthropic’s own research found that Claude 3.7 Sonnet “mentioned the hint 25% of the time” when a hint shaped its answer, so a model’s narration can’t serve as the record of a decision.
Practitioners such as Tian Pan have made this argument generally; Anthropic’s documentation shows it in one product. It extends Complete Decision Cycle Capture, one of Futurum’s Seven Principles of Observability-Native: the record of why comes from the application, not the model’s narration.
The reason has to be written down as part of the work: the decision, the policy applied, and the evidence used, in the custom tools and application code builders’ control. Without that record, teams reconstruct intent in the middle of an incident or an audit.
Anthropic should give Managed Agents spans a documented field for builder-supplied decision records and a redacted content option for restricted customers before the export leaves beta.
Anthropic Sends Traces to Incumbents and Keeps Its Own Grader
AWS Bedrock AgentCore sends traces to Amazon CloudWatch, Google’s Agent Runtime to Cloud Trace, and OpenAI’s Agents SDK to OpenAI’s backend by default.
Among model providers and cloud agent runtimes, Claude agent observability differs in where it sends its spans. Anthropic sends the spans its own servers produce to the tool the customer already runs, with no customer code, and works with incumbent observability vendors instead of replacing them.
New Relic launched Claude Platform tracing on October 14, arguing that “because Claude’s spans share a trace with your APM data, you can follow the trail past the agent, which LLM-only tools can’t do.” Dynatrace, absent from Anthropic’s partner list, published a Claude Platform setup guide on October 16, two weeks after completing its acquisition of Arize, which is on the list. Splunk‘s Claude integration is in public beta.
Evaluation is where the incumbents are investing, and Anthropic emits evaluation spans of its own. Its anthropic.session.outcome\_evaluation span carries the result and, in Full content mode only, “the grader’s explanation of the result, as plain text.” That is rubric scoring with a written explanation, one slice of what evaluation vendors sell, shipped inside the traces Anthropic sends to them. Part of the grading layer now sits in the runtime Anthropic hosts, while the open GenAI semantic conventions keep the spans readable by any tool that supports them.
Buyers can test which platform fits with one failed agent task. Ask the current observability vendor and any evaluation specialist to follow it across the model, the tools, and the application’s own services, name the missing evidence, and turn it into a repeatable evaluation. A platform that can’t do that with Claude’s spans hasn’t earned the agent budget.
Buyers Should Treat the Audit Record as a Sourcing Decision
For the CIO, the CFO, and internal audit, the choice between Managed Agents and the Messages API decides who controls the evidence. Letting Anthropic host the runtime takes orchestration work off the engineering backlog. It also hands Anthropic the call on what each turn records, so the buyer’s control framework now depends on a span reference the buyer doesn’t write.
Regulated buyers get the least evidence where they need the most. A HIPAA, FERPA, or K-12 restriction cuts spans from Anthropic’s servers to metadata, and Managed Agents isn’t eligible for zero data retention. A healthcare or education organization running a Managed Agents workflow gets a timeline of tool calls with no content, from a runtime whose retention it doesn’t set. Running the Messages API with its own orchestration and SDK tracing keeps content in the buyer’s hands, at the cost of building and operating the agent loop. The audit requirement should make that buy-versus-build call before convenience makes it for them.
The decision record is also a budget line someone has to own. Writing down the decision, the policy applied, and the evidence used is application code, and it lands on product engineering while the traces come out of the observability budget. Full content mode adds its own cost. Copying prompts and responses into the observability platform raises ingest volume and creates a new store of sensitive data that needs access controls and a retention policy of its own.
Most organizations are making this call from a weak base. In Futurum’s 2H 2026 SLE Decision Maker Survey, 32% gate irreversible agent actions behind human approval, AI agent governance ranks as the least mature practice in the SLE portfolio, and 42% name missing governance infrastructure among the top limits on expanding AI use. What organizations can observe, control, and prove sets how fast agents reach regulated work, more than what the models can do.
Ownership has to be settled before a Console Admin adds the first destination. Internal audit or compliance defines what a decision record must contain for each agent workflow. The CISO or platform lead sets the content-mode policy and names the system that holds the audit record, since exports can drop spans. Engineering builds to that specification. Procurement then puts three questions to Anthropic, or to any hosted agent runtime: what each span carries in the buyer’s content mode, what happens to a span that fails delivery, and how much notice the buyer gets when span names and attributes change. On the last one, Anthropic’s documentation is plain for now: during the beta, they can change.
What to Watch:
- Whether Anthropic documents a field for builder-supplied decision records on Managed Agents spans before the export leaves beta.
- Whether HIPAA, FERPA, and K-12 customers get a redacted content option by mid-2027.
- Whether OpenAI, Google, or AWS allow an admin to send their hosted runtimes’ own spans to third-party tools via a console setting in the first half of 2027.
- Whether observability vendors ingest Anthropic’s outcome-evaluation spans as eval input, or score Claude agents only with their own graders.
- Whether Anthropic adds delivery guarantees or a replay option for API export before it leaves beta.
See Anthropic’s October 2026 post on OpenTelemetry tracing on the Claude website.
Sources
- Anthropic Claude Agent Observability
- With Kong Volcano, Kong Hosts the Agents It Governs, Futurum
- OpenAI Moves Up the Stack and Competes With the Platforms It Powers, Futurum
- Claude Fable 5 Is Most Consequential Where Software Is Built, Futurum
- Splunk .conf26: Trust is Key for the Agentic Era, Futurum
Declaration of generative AI and AI-assisted technologies in the writing process: This content has been generated with the support of artificial intelligence technologies. Due to the fast pace of content creation and the continuous evolution of data and information, The Futurum Group and its analysts strive to ensure the accuracy and factual integrity of the information presented. However, the opinions and interpretations expressed in this content reflect those of the individual author/analyst. The Futurum Group makes no guarantees regarding the completeness, accuracy, or reliability of any information contained herein. Readers are encouraged to verify facts independently and consult relevant sources for further clarification.
Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.
Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.
Read the full Futurum Group Disclosure.
Other Insights From Futurum:
WEKA and Smartbird End the Sovereignty-Performance Tradeoff
Sword Group’s 17% Organic Growth Validates AI Transformation Bet
Western Digital’s 55% Revenue Surge Signals AI Storage Boom
Author Information
Vikram Rathnam is Research Director, Software Lifecycle Engineering at The Futurum Group. His research examines how software is built, tested, secured, and operated as AI agents move from assisting developers to doing the work. His coverage includes observability, platform engineering, and the control planes that govern agents. Before joining Futurum, he spent 24 years on the vendor side: ten in engineering at Silicon Labs and fourteen in product and partner strategy, with roles at Cisco, Secureworks, Dell, and CMIT Solutions. Across those roles, he managed B2B cybersecurity, IoT, and AI products, with partner and channel ecosystems as the common thread. At CMIT Solutions, he led product strategy across a franchise network of 180 locations. He reads a vendor’s roadmap as someone who has had to ship one and take it to market.
Mitch Ashley is VP and Practice Lead for the CIO & Technology Buyers and Software Lifecycle Engineering practices at The Futurum Group. A multi-time CIO and CTO with 30+ years leading technical organizations, Mitch built and operated production systems spanning cybersecurity for the U.S. Department of Defense, PKI services for the broadband and 5G industries, SaaS platforms, large-scale telecom and banking systems, and a national broadband network. His work with AI began early, developing expert systems that diagnosed and repaired complex mainframe environments. That operator foundation grounds his analysis in operational consequence, covering the technology buyer's world of software engineering, cybersecurity, DevOps, cloud, and AI.
