OpenTelemetry’s CNCF Graduation Turns AI Agent Observability Into a Standards Problem

The Cloud Native Computing Foundation has graduated OpenTelemetry to sit alongside Kubernetes and Prometheus as a stable, production-ready standard. The project’s own maintainers say the work that follows, standardizing observability for AI agents, is far from finished.

CNCF announced OpenTelemetry’s graduation on May 21, 2026, and the OpenTelemetry community’s own post confirms the same date. Two months later, Adriana Villela of Dynatrace and Reese Lee of New Relic used a CNCF blog post to lay out what the milestone actually certifies and where the project goes from here. For platform and security teams still treating OpenTelemetry as an optional instrumentation choice, both posts make the same point: that decision is no longer really open for debate.

What CNCF’s Graduated Tier Actually Certifies

CNCF grades hosted projects across three maturity levels. Sandbox covers experimental work not yet widely tested in production. Incubating covers projects already running successfully in production for a smaller set of users with a healthy contributor base. Graduated is reserved for projects CNCF considers stable, widely adopted, and production ready, with thousands of contributors behind them, the same tier Kubernetes and Prometheus occupy. Moving between tiers requires a vote from CNCF’s Technical Oversight Committee, not a maintainer’s self-assessment.

Per the CNCF blog post from Villela and Lee, a project has to clear four separate bars to graduate:

• Production adoption. Multiple independent organizations have to be running the project in production; the post names GitHub and Farfetch as examples for OpenTelemetry specifically.
• Robust governance. A documented governance model with clearly defined roles for election and retirement, plus transparent decision-making.
• Community health. An established PR review process and a base of regular contributors spread across multiple organizations, with reviewers that respond in a timely manner.
• Security. At least one independent security audit, with critical findings addressed.

OpenTelemetry cleared all four having started its formal path to graduation in 2025 and reaching Graduated status in May 2026. According to the same CNCF post, it is now the second-highest velocity project in the entire foundation, trailing only Kubernetes, with more than 12,000 contributions from over 2,800 companies and hundreds of maintainers spread across language-specific Special Interest Groups.

The Fragmentation Problem OpenTelemetry Was Built to Close

The CNCF post is candid about why any of this mattered in the first place. Before OpenTelemetry, telemetry signals were not standardized: every vendor built and maintained its own instrumentation libraries, so switching providers meant physically ripping one vendor’s code out of an application and replacing it with another’s. Traces, logs, and metrics were also treated as separate, disconnected signals, which made correlating them across a request’s full path a manual, error-prone exercise rather than something the tooling did for you.

Two earlier projects tried to fix this before OpenTelemetry existed: OpenTracing, hosted at CNCF, and OpenCensus, built by Google. Rather than compete indefinitely, the two merged into what became OpenTelemetry. Their predecessors were formally retired once the merger held: OpenTracing was archived in January 2022, and OpenCensus followed in July 2023. That consolidation, backed by every major observability vendor plus an active end-user community, is what let OpenTelemetry become what the CNCF post calls the de facto open standard for telemetry, a claim CNCF’s own announcement post title repeats almost verbatim.

Adoption in Numbers, Not Just Vendor Endorsements

CNCF’s graduation announcement backs the “de facto standard” framing with usage data rather than just adopter logos. Over the past twelve months, the OpenTelemetry JavaScript API package was downloaded more than 1.36 billion times, and the Python API package passed 1.3 billion downloads, with both packages setting new monthly download records in April 2026. The announcement names Alibaba, Anthropic, Bloomberg, Capital One, eBay, FICO Software, and Heroku among the organizations relying on OpenTelemetry to monitor and secure their systems, and notes recently added support for languages including Kotlin, alongside Profiles, a newer telemetry signal that is now officially in alpha.

What Graduation Changes for Teams Still on the Fence

The practical argument in the CNCF post is aimed squarely at holdouts. Dan Gomez Blanco, a maintainer of the OpenTelemetry End User Special Interest Group, is quoted from a LinkedIn post making the case directly: “For end users, this graduation signals that OTel is far from being an ‘emerging standard’. Its contributor health, security and quality standards, governance processes, and wide adoption have been evaluated to be at the level required by any enterprise, of any scale.”

That framing matters operationally. Graduated status is a CNCF-audited answer to the questions a platform team’s own architecture review would otherwise have to answer itself: is this project’s governance stable enough to depend on, has its security posture been independently checked, and will its contributor base still be active in three years. For teams evaluating an observability vendor or platform today, that due diligence on the instrumentation layer itself is effectively already done. The remaining decision is which backend, dashboarding, or analysis tool to pair with it, not whether OpenTelemetry’s data model can be trusted.

Peco Karayanev, CTO at Autoptic, framed the same shift from the buyer’s side in CNCF’s announcement: “OpenTelemetry’s graduation is a big step for teams like ours building automation and AI-driven operational workflows. Having more consistent telemetry standards across tools makes it easier to connect systems, reduce integration work and use operational data more effectively across environments. It also creates a stronger foundation for sharing context between teams, platforms, workflows and AI systems without being tied to a single vendor stack.”

The Next Frontier: Observability Built for AI Agents

Both CNCF posts are explicit that graduation is a milestone, not an end state. The OpenTelemetry community’s own announcement puts it plainly: “Graduation is not the finish line.” CNCF’s announcement ties that unfinished work directly to AI: the project’s adoption, it says, is “gaining new interest as a layer to observe performance, reliability, accuracy and trustworthiness in AI workloads,” a use case that barely existed when OpenTelemetry’s early API design was being settled.

The Villela and Lee post is more specific about where that work is headed. OpenTelemetry’s maintainers see a strong need for observability around new categories of workloads, agentic workflows chief among them, an area already being addressed through emerging generative AI semantic conventions rather than left for every vendor to define independently. Two areas the project had not prioritized as heavily until now, browser and mobile observability, are also called out as active gaps. On the tooling side, the post highlights three projects aimed at making OpenTelemetry easier to operate at scale rather than just easier to adopt: OpenTelemetry Weaver, which helps teams define and govern their telemetry schemas as instrumentation sprawls across services; OpenTelemetry Packaging, which bundles components into installable modules; and the OpenTelemetry Injector, aimed at zero-code instrumentation for services that were never instrumented to begin with.

For teams already standing up monitoring around AI agents and LLM-backed workflows, that roadmap is the more actionable signal than the graduation vote itself. Building a one-off logging scheme for agent traces now risks the same rework that hit teams who built deep on proprietary instrumentation libraries before OpenTelemetry consolidated the market. Teams evaluating how to observe agentic workloads today have a narrower, better-informed choice than teams evaluating general telemetry did a decade ago: track the semantic conventions work directly, rather than build a bespoke standard that graduation has already made a harder sell.