---
title: "Agent Observability: Monitoring Decisions, Not Requests"
description: "Agent observability captures the sequence of decisions an agent makes between model calls, not just request-level metrics, and is distinct from monitoring (health) and evaluation (quality). Unlike deterministic database systems, an agent can succeed at every step yet fail overall—for example, looping forever on a failed fix while spans report success. The primary artifact is a trace, a parent-child tree of spans covering agent invocations, model calls, tool executions, and memory reads and writes. OpenTelemetry GenAI conventions define span types like create_agent and execute_tool, but the standards are still in development and change frequently; vendor support includes Datadog, Honeycomb, and frameworks like LangChain and CrewAI. A reasoning trace layer on top of operational telemetry explains why an agent chose that path; in multi-agent systems the trace must span the full coordination chain, not one agent’s output."
---

# Agent Observability: Monitoring Decisions, Not Requests

[MongoDB](https://yomu.fyi/company/mongodb) · Aug 25, 2026

**Type:** Explainer

## Summary

Agent observability captures the sequence of decisions an agent makes between model calls, not just request-level metrics, and is distinct from monitoring (health) and evaluation (quality). Unlike deterministic database systems, an agent can succeed at every step yet fail overall—for example, looping forever on a failed fix while spans report success. The primary artifact is a trace, a parent-child tree of spans covering agent invocations, model calls, tool executions, and memory reads and writes. OpenTelemetry GenAI conventions define span types like create\_agent and execute\_tool, but the standards are still in development and change frequently; vendor support includes Datadog, Honeycomb, and frameworks like LangChain and CrewAI. A reasoning trace layer on top of operational telemetry explains why an agent chose that path; in multi-agent systems the trace must span the full coordination chain, not one agent’s output.

## Context

The post is the sixth part of a series on designing an agentic platform, written by MongoDB engineers. It addresses the gap between traditional request-level monitoring and the observability needed for non-deterministic agent systems.

## Approach / What changed

The approach is to instrument agent systems using OpenTelemetry GenAI semantic conventions for traces, capturing spans for agent invocations, tool executions, and memory access. Additionally, a reasoning trace layer captures the sequence of decisions, and in multi-agent systems the trace spans the full coordination chain.

## Takeaways

- Agent observability isolates three distinct jobs: monitoring (system health), observability (reconstructing what the agent did and why), and evaluation (judging quality).
- An agent can succeed at every individual step yet fail overall, as in a loop that never converges while individual spans report success; traditional request-level monitoring cannot detect this failure class.
- OpenTelemetry GenAI semantic conventions define agent span types (create\_agent, execute\_tool) and are supported by vendors like Datadog and frameworks like LangChain; spans must carry attributes beyond name and duration to enable cross-trace queries.

## FAQ

### How does agent observability differ from database monitoring?

Database monitoring is for deterministic systems where failure is explicit (slow query, missing index). Agent systems are non-deterministic; every step can succeed while the overall run fails, so observability must capture decision sequences, not just request metrics.

### What is the recommended standard for agent traces?

The OpenTelemetry GenAI semantic conventions define span types like create\_agent, invoke\_agent, execute\_tool. As of mid-2026 they are still in development, have changed frequently, but are supported by Datadog, Honeycomb, New Relic, Grafana, and frameworks like LangChain and CrewAI.

### Why must memory access be logged in the trace?

Memory access (queries and returned context) must appear in the trace next to model and tool spans so that a failure can be attributed to retrieval issues rather than being mistaken for reasoning or tool problems.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [Observability](https://yomu.fyi/topic/observability), [OpenTelemetry](https://yomu.fyi/topic/opentelemetry)

- Source: [MongoDB](https://www.mongodb.com/company/blog/technical/agent-observability-monitoring-decisions-not-requests)
- Source URL: https://www.mongodb.com/company/blog/technical/agent-observability-monitoring-decisions-not-requests
- Ingested by Yomu: 2026-09-20T07:52:04.027Z

[Read original post](https://www.mongodb.com/company/blog/technical/agent-observability-monitoring-decisions-not-requests)