AI observability with MCP and OpenTelemetry: how the design works

Updated on October 3, 2026

A 2025 three-layer system fed context-enriched telemetry into an MCP server for AI-driven analysis; how it works and what changed.

En 30 segundos Modern systems emit logs, metrics and traces that rarely line up. A 2025 design proposed fixing that in three layers: enrich telemetry at creation with OpenTelemetry, expose it through an MCP server, and let an AI engine analyse it. Since then, the 2026-07-28 MCP specification retired the initialize handshake and the session header, and OpenTelemetry moved its GenAI conventions to a new repository.

The problem: signals that do not connect

In August 2025, VentureBeat published an article by Pronnoy Goswami describing e-commerce platforms that produce tens of terabytes of logs per day, plus millions of distributed traces and thousands of correlation IDs. The article cites New Relic's 2023 Observability Forecast: 50% of organizations report siloed telemetry, and only 33% have a unified view of metrics, logs and traces.

Observability means being able to understand a running system from the data it emits. When logs, metrics and traces live apart, engineers have to stitch them together by hand during an incident, which slows detection and resolution. The article frames the payoff as a reduced MTTD and MTTR, which it spells out as minimum time to detect and minimum time to resolve.

The three-layer design

The first layer is context-enriched telemetry generation. Standardised metadata is attached to every signal when it is created, using OpenTelemetry to generate traces, so correlation is solved at the source rather than afterwards.

The second layer is an MCP server. The Model Context Protocol is an open protocol that lets AI systems fetch data from tools in a structured way. Here it acts as a data pipeline: raw telemetry becomes something an AI can query, filter and aggregate. The third layer is the analysis engine, which consumes data through that interface and applies statistical methods such as z-scores to flag anomalies and look for root causes.

Our simplified view of the architecture described in the 2025 article.

What happened next

Two of the pieces have moved. On 28 July 2026 the MCP maintainers published a new specification that retired the initialize/initialized exchange along with the Mcp-Session-Id header, so each request travels on its own and any request can land on any server instance behind a plain round-robin load balancer.

The specification's versioning page now names 2026-07-28 as the current protocol version and describes revisions 2025-11-25 and earlier as handshake-based. Clients that want to choose a version up front can call server/discover, which returns the server's supported protocol versions, capabilities and identity. An MCP server built as the 2025 article describes would have used the handshake model that this revision replaces.

On the telemetry side, the OpenTelemetry GenAI conventions now live in a dedicated repository. A May 2026 OpenTelemetry post says they are in use but still under active development. It lists standard attributes such as gen_ai.request.model and token counts, and notes that prompt content and tool arguments are not captured by default because they can hold sensitive data; capturing them is opt-in.

¿Me afecta? If you run complex services, the core idea holds: attach consistent context to signals at creation, then let tools or AI query them. If you build on MCP, check the 2026-07-28 changes before reusing 2025-era server designs. If you instrument LLM calls, expect the GenAI conventions to keep changing and decide deliberately whether to capture prompt content.

Opinion · Open question

The 2025 design leaned on MCP as its data pipeline, and the 2026-07-28 MCP specification then retired the initialize handshake and the session header that design relied on. Meanwhile OpenTelemetry's GenAI conventions are still described as under active development.

Open question: how much of any observability stack built today will survive the next revision of the standards it rests on? The sources show the pace of change but no answer, and teams that wire AI into incident response need one before they commit.

Related articles from LaRebelión

OpenAI probes more possible AI hacking incidents after the Hugging Face breachApple tightens full-disk access to rein in AI agentsOpenAI Reports Rogue AI Agents Affecting Over 100 Organizations

Original source: VentureBeat

Produced with AI support and reviewed by the newsroom

Byline · Chief editor · English edition · London

“Fix the correlation problem where signals are born, then let the AI read; the plumbing underneath keeps moving.”