,

MCP Helps Agents Act. OpenTelemetry Helps Us See What They Did.

Illustration paired with Jason Fishbein’s LinkedIn post about MCP, OpenTelemetry, and agent observability.

By now, you have probably heard of MCP. But have you heard of OTel?

Model Context Protocol helps AI agents connect to tools, systems, and data. It is one of the reasons agents can move beyond chat and start doing useful work in the business.

But there is another standard leaders should understand as those agents become more autonomous: OpenTelemetry, or OTel.

Think of it this way: MCP helps agents do things. OTel helps us see what they are doing.

What OTel adds

OpenTelemetry has been used since 2019 to instrument applications and distributed systems. It provides a common way to collect operational evidence instead of leaving each system with its own disconnected logs.

That matters for agents because one business task can cross a model, an orchestration layer, several tools, one or more MCP servers, and an approval workflow. When something goes wrong, a green “running” indicator is not enough.

Well-instrumented agent workflows can capture evidence such as:

  • which model and configuration handled a step;
  • which tools or MCP servers the agent called;
  • timing, token consumption, failures, and retries;
  • how work moved between agents, tools, and workflows; and
  • where a task ended, escalated, or failed.

The emerging GenAI and agent-related OTel conventions help teams express more of that information consistently. That gives engineering, security, and operations teams a better chance of following one assignment across systems rather than reconstructing it from a pile of vendor-specific dashboards.

A standard is not the same as complete visibility

This is the part that gets missed. Standardized observability does not guarantee complete observability.

One implementation may record model calls, duration, and token use. Another may expose rich traces of tool inputs, policy checks, and execution steps. Both can use OTel. Neither automatically tells you whether the agent had all relevant context, made the right decision, or was authorized to act.

Instrumentation is evidence. It is not judgment.

That distinction matters as agents operate faster and across more systems than a person can watch directly. Human review still belongs where judgment is required. But at scale, teams also need automated monitoring and evaluation capable of identifying unusual behavior, missing approvals, repeated failures, or decisions that drift from the expected business outcome.

Observability does not enforce the boundary

OTel can help a team detect the signals leading up to an incident, understand what occurred, and respond faster afterward. It does not stop an agent from exceeding an authority boundary by itself.

That requires enforcement where the action happens: scoped permissions, approval gates, transaction limits, policy checks, and the ability to revoke access. A trace that shows an unauthorized refund is useful. A system that rejects it before it happens is better.

The practical question for leaders is not simply whether an agent platform has monitoring. Ask what evidence one completed assignment produces, which gaps remain, who reviews alerts, and which controls prevent prohibited actions rather than just recording them.

Scale action and oversight together

MCP is helping organizations scale what agents can do. OTel could become equally important in scaling how organizations monitor that work.

You cannot monitor what you cannot observe. And you cannot control what you do not understand.

Trust in autonomous agents will not come from what we hope they will do. It will come from our ability to observe, verify, and control what they actually do.