OpenTelemetry Observability for Builders

Every governance capability covered earlier in this series, threat detection, compliance monitoring, the Agents-at-Risk model, quietly assumes that agent activity data already exists somewhere for those systems to read. This article is about how that data actually gets created in the first place, from a builder’s side of the wire.

1 · Visibility is not automatic; someone wires it up

An agent does not become observable by existing. Someone has to instrument it, meaning add the code or configuration that captures what the agent does and sends that record somewhere it can be watched. This article is about the choices a builder has for doing that instrumentation, and how much work each choice actually requires.

2 · OpenTelemetry: one open standard, not a Microsoft invention

Agent 365 observability builds on OpenTelemetry, often shortened to OTel, an open, industry-wide standard for capturing telemetry, and it provides “a unified framework for capturing telemetry consistently and securely across all agent platforms.” The word “open” matters: this is not a proprietary Microsoft format that locks a builder in. It is a shared standard, which is precisely why the same tracing approach can work whether an agent was built with Microsoft’s own tools or with a completely different vendor’s stack.

3 · The Microsoft OpenTelemetry Distro

To make adopting that open standard simpler, Microsoft offers the Microsoft OpenTelemetry Distro: a single observability toolkit that provides “a single observability SDK across Microsoft, powering Agent 365, Microsoft Foundry, Azure Monitor, and more.” Rather than learning a separate tracing approach for each Microsoft product an agent touches, a builder learns one distribution that already speaks to all of them. An AI-guided setup can even configure this automatically in many cases, further lowering the effort required to get started.

4 · Auto-instrumentation: the easy path

The simplest route is auto-instrumentation: if an agent is built using OpenAI’s tools, LangChain, or Microsoft’s own Agent Framework, tracing can be enabled with very little extra work, because the integration for those popular stacks already exists. A builder working in one of these ecosystems gets meaningful observability largely for free, without hand-writing the tracing calls themselves.

5 · Manual instrumentation and direct OTLP

For everything else, manual instrumentation means the builder writes the tracing calls into their own code by hand, choosing exactly what gets recorded. And for a further edge case, if an agent’s framework cannot use the Agent 365 SDK at all, or the agent is written in a language the SDK does not yet support (Java is named as an example), there is still a path in: sending telemetry directly to Agent 365 over raw OTLP (the OpenTelemetry transport protocol) via HTTP, with no SDK dependency whatsoever. This means observability is never fully out of reach, even for a stack the official tooling has not caught up to yet.

6 · Super agents: logging the delegation chain

Some agents are what Microsoft calls a super agent: one that internally calls one or more sub-agents to complete a task. When that happens, those internal sub-agent calls need to be logged too, either as an ordinary tool call or as an agent-to-agent (A2A) call, whichever fits, so the full chain of delegation stays visible and auditable. Without this, a super agent’s internal handoffs would vanish into a black box the moment work passed from one agent to another, exactly the kind of unseen connection this whole series has warned against at every level.

7 · Glossary — every short-form term, spelled out

Instrumentation

Adding code or configuration to a system so its activity can be captured and observed.

OpenTelemetry (OTel)

An open, industry-wide standard for capturing telemetry consistently across many platforms.

Microsoft OpenTelemetry Distro

A single Microsoft-provided observability toolkit powering Agent 365, Microsoft Foundry, and Azure Monitor together.

Auto-instrumentation

Tracing enabled with minimal extra work because integration already exists for a popular framework, such as OpenAI, LangChain, or Agent Framework.

Manual instrumentation

Tracing calls written by hand into an agent’s own code for stacks without ready-made integration.

OTLP (OpenTelemetry Protocol)

The transport protocol OpenTelemetry uses to send telemetry data, usable directly without any SDK dependency.

Super agent

An agent that internally calls one or more sub-agents to complete a task.

Agent-to-agent (A2A) call

A logged interaction between two agents, used to keep a delegation chain auditable.