The migration nobody wanted to do
The renewal quote arrives. It is higher than last year. Someone asks: why not switch?
The answer is rarely the licence. It is the code. Hundreds of services carry one vendor’s agent. Leaving means re-instrumenting the estate and running two stacks while you do it. So the contract gets renewed.
Now take the hardest version of the same problem. A regulated enterprise runs two OpenShift clusters and 250+ virtual machines. The network is air-gapped. Application images cannot be changed. No data can leave. Here, the vendor model does not just lock you in. It locks you out. The standard way to add an agent is to rebuild the image, and that is not allowed.
Both stories have the same root. The vendor owned the instrumentation. OpenTelemetry moved that ownership back to the enterprise. This article shows what that changes, and returns to that air-gapped estate at the end.
What changed: the stack came apart
For years, observability came as a bundle. The vendor’s agent, data format and backend were fused. Your data, their schema.
OpenTelemetry separated those layers. Instrumentation is written once, against open SDKs. Data travels over one open protocol, OTLP. The OpenTelemetry Collector, run by the enterprise, shapes and routes it. Any OTLP-compatible backend can receive it. Instrumentation itself becomes something the enterprise controls. It can even be injected at pod startup, with no change to the application image.
In May 2026, the CNCF moved OpenTelemetry to graduated status, its highest maturity level (CNCF announcement). For enterprise buyers, that settles whether it is here to stay.
On the left, every layer belongs to one vendor. On the right, the enterprise owns the instrumentation and the pipeline, and backends become interchangeable.
Before and after
Every row moves a decision from the vendor to the enterprise.
What it changes for the enterprise
- Backend freedom. Pick tools per use case, not per contract. Change them without touching application code. That alone changes renewal conversations.
- The Collector as control point. Filter, enrich, sample, mask and route in one governed pipeline. For regulated enterprises, data residency and PII rules are enforced before data leaves the environment. Or, in an air-gapped estate, data never leaves at all.
- Cost by design. Teams decide what is stored at full fidelity and what is sampled. You stop paying to store data nobody queries.
- One language across teams. Semantic conventions give developers, SREs and security teams the same identifiers for the same incident.
Why it matters more with AI
LLM apps and agents chain model calls, retrievals and tool calls. Without tracing, a slow or wrong answer is a black box. OpenTelemetry’s GenAI semantic conventions, covering agents and MCP, address this. They are still in development, so expect field names to change.
The reverse holds too. AI-assisted analysis is only as good as the telemetry beneath it. Open, correlated data is part of the foundation for running AI in production.
The honest part
Freedom comes with responsibility. Control has to be exercised.
- The Collector is production infrastructure. It needs sizing, high availability and upgrades.
- Maturity varies. Traces, metrics and logs are established. Profiles entered public alpha in March 2026 and are not yet meant for critical workloads.
- Legacy and restricted estates need design. Hybrid models are normal, as the example below shows.
- Conventions need governance. Without naming standards, two teams produce data that does not correlate.
In practice: back to the air-gapped estate
Return to the enterprise from the opening. Its question was never which backend to buy. It was whether it could instrument its own estate, on its own terms. One more constraint made it harder: cluster-admin rights for the standard OTel Operator webhook were not available. Ashnik’s design used OpenTelemetry to work within every one of them.
- No code or image changes. Namespace-scoped auto-instrumentation injected the OTel Java agent. One label and one annotation per application.
- Pipeline under enterprise control. An in-cluster Collector, with no external endpoint.
- Open protocol to the backend. Traces went over OTLP to Elastic APM, alongside logs and metrics.
The result was request-level tracing across the Java estate where none existed, with zero image changes. Every layer, from instrumentation to pipeline, stayed under the enterprise’s control. Logs and metrics used OpenShift’s Cluster Log Forwarder and Metricbeat. Hybrid, by choice (full walkthrough).
This is how Ashnik approaches OpenTelemetry: as part of an Observability & Search Platform, built through Assess → Architect → Implement → Modernize → Optimize → Operate, with right-fit technologies such as Elastic, Grafana, Netdata and Cardinal.
Five questions for technology leaders
- Do we own our telemetry, or does our vendor?
- What would it cost us to change backends tomorrow?
- Are we paying to store data nobody queries?
- Who controls our telemetry pipeline?
- Is our observability ready for the AI workloads we plan to run?
The shift from vendor agents to enterprise assets is not a product upgrade. It is a decision to take control back.
Talk to Ashnik about assessing your observability estate.
