CloudWatch Omni Bets on OpenTelemetry's Unstable GenAI Spec

A new AWS product on a moving foundation

Amazon has introduced CloudWatch Omni, which it describes as a unified observability experience for application and AI workloads. AWS characterizes it as app-centric, AI-powered, built on open standards, and delivered outside the main console 5. The pitch is aimed at teams building AI agents. Omni promises tracing, evaluation, and experimentation across any model provider, framework, or runtime. It builds on an "eval-driven workflow," with built-in evaluators for quality, correctness, and coherence. Developers reach it from their IDE or a standalone web app rather than the AWS Management Console 5.

The launch matters less for any single feature than for its timing. AWS is making open standards central to an agent-observability product at a moment when the most likely candidate for that standard, OpenTelemetry's generative-AI semantic conventions, has not reached stable status.

What the OpenTelemetry GenAI conventions are

The conventions are a shared vocabulary of span, metric, and attribute names for instrumenting generative-AI and agent systems 1. They cover single inference calls, embeddings, tool execution, retrieval, memory, and complete agent runs 3. Operations such as invoke_agent and invoke_workflow carry attributes like gen_ai.agent.name and gen_ai.workflow.name. That lets a backend recognize an agent run or an orchestrated multi-step process regardless of which framework produced it 3. The spec also defines Model Context Protocol tracing and required metrics for latency and token usage 12.

The value proposition is consistency. If an LLM call, an agent invocation, and a tool execution appear in a trace under the same attribute names no matter the model, framework, or backend, teams can switch vendors without rewriting instrumentation 4. For site reliability engineers already used to traces and spans, agents become one more instrumented system with a defined vocabulary for planning, delegation, and tool use 1.

"Development" is not a formality

All the coverage agrees on one point: the conventions are not stable. Every agent, tool, and MCP span is marked "Development" 1. Nearly all gen_ai.* attributes carry that badge, which means names can change without a major version bump 2. One practitioner guide states plainly that the conventions remain pre-stable and experimental, with no 1.0 release 4.

The most notable recent change was structural. In semantic-conventions v1.42.0, released 12 June 2026, the GenAI definitions moved out of the core OpenTelemetry repository into a dedicated semantic-conventions-genai repo 34. The goal was to let the area iterate faster than the core project's stability bar allows 3. Reading the split as maturity would be a mistake. One analysis calls it organizational rather than a stability promise 4. Some guides still reference v1.41 as the current baseline 2, a small sign of how quickly the documentation landscape shifts.

Why the market isn't waiting

If the standard is unfinished, why are vendors building on it now? Demand is one reason. One estimate puts the LLM observability market at roughly $2.69 billion in 2026, growing around 36% annually. The same source flags that figure as directional 2. It also cites Gartner data suggesting only about 15% of generative-AI deployments were instrumented in early 2026, leaving most effectively blind 2. That gap is a large commercial opening, and a vendor that waits for a 1.0 spec risks ceding it.

The competitive field is also crowded. Self-hosted tools, managed SDKs, and proxy gateways all compete for agent telemetry 2. Aligning with an emerging open vocabulary is how a newcomer signals it won't lock customers in.

The reading: useful, but treat "open standards" with care

AWS's announcement emphasizes open standards but does not detail which convention versions Omni implements 5. It is reasonable to infer that the OpenTelemetry GenAI work is the relevant reference point, since it is the only broadly recognized vendor-neutral vocabulary for this telemetry. But that is an inference, not something AWS has spelled out in its launch materials.

The practical consequence for buyers is that "built on open standards" currently means "built on a draft." Attribute names in dashboards, alerts, and evaluation pipelines may need migration as the conventions evolve. The faster cadence enabled by the June repo split 3 could make those changes more frequent in the near term, not less. Teams adopting Omni, or any competitor, should ask which convention version is supported and how the vendor handles renames. They should also avoid hard-coding queries against attributes that could change.

None of this makes Omni a bad bet. Betting on a shared, if unfinished, vocabulary is better than proprietary schemas, and AWS's adoption could speed the spec toward stability. The honest framing is that Omni rides a standard still being written.