AI Observability: Monitoring AI Agents Across LLM, Tool, and File Activity · Highflame
Highflame Identity is now open source: agent identity on open standards. Read the launch
Reference

AI Observability

AI observability is the practice of capturing and querying everything your AI agents actually do at runtime, not just their model calls. It spans LLM prompts and completions, MCP tool calls, file reads and writes, and network activity, across code agents, web agents, and custom in-house agents, all tied to the identity that took each action.

AI observability vs LLM observability

Most tools sold as observability for AI are really LLM observability: they capture the model call, the prompt, the completion, the token count, and the latency. That is useful, and it is where the category started, with products built around prompt logging and cost dashboards.

But an agent does far more than call a model. It runs shell commands, reads and writes files, calls tools over MCP, hits the network, and spawns sub-agents. AI observability is the broader practice of seeing all of it, not just the model turn in the middle. If your observability stops at the LLM call, you can see what the agent thought and none of what it actually did.

Logging vs monitoring vs observability

The three words answer different questions, and the distinction decides whether your setup can actually explain an incident.

LoggingMonitoringObservability
Question it answersWhat happened?Is it healthy right now?Why did this happen, and who did it?
The dataDiscrete events, line by lineMetrics and alerts on known signalsCorrelated traces across a whole session
For AI agentsA record of each callDashboards of latency, error rate, and costQuery any agent’s full session, its tools, files, and delegation

The full picture: every action an agent takes

AI observability means capturing the whole surface of what an agent does:

  • LLM calls. Prompts, completions, tokens, and latency (time to first token and total).
  • Tool and MCP calls. Which tool, with which arguments, returning what.
  • File activity. Reads and writes, which matter enormously for a coding agent with disk access.
  • Network and web requests. Where the agent reached, and what it sent or pulled back.
  • Delegation. Which sub-agents it spawned, and how deep the chain went.

And the signal that ties it together, the one plain LLM logging never captures, is identity: which agent took each action, who owns it, and whether policy allowed it. Without that, you have a pile of activity and no way to say who is behind it.

Observability across every kind of agent

AI observability is not one integration, because agents run in very different places. It has to reach all of them.

Code agent observability

Coding agents like Cursor, Claude Code, and GitHub Copilot read repositories, run shell commands, edit files, and call MCP servers from inside the IDE and the CLI. Observing them means capturing those commands, file changes, and tool calls in the execution path, not just the model prompt, so you can see what a coding agent actually did to a machine.

Web AI observability

Agents that browse and act on the web navigate pages, fill forms, and pull data. The principle is identical: capture the actions the agent took and the data that moved, and tie each to the agent’s identity, so a web-acting agent is as accountable as one calling an internal API.

Custom agent observability

Most enterprises also run bespoke agents built on their own stack. These instrument cleanly with OpenTelemetry, emitting the same kind of trace data as everything else, so a custom agent’s LLM calls, tool calls, and actions flow into the same surface rather than a separate silo.

AI observability vs APM

Application performance monitoring (Datadog, New Relic, and the like) was built for a service: a request comes in, spans fan out across microservices, a response goes back. An agent is not a request. It is a reasoning loop that calls tools, touches files, and spawns sub-agents over a long session.

APM will show you the HTTP spans underneath an agent. What it cannot show you is “this agent, two hops into a delegation chain, acting for a deactivated user, wrote to this file and called the payments tool.” The gap is agent behavior and identity, and it is exactly where incidents become impossible to reconstruct.

OpenTelemetry for AI observability

OpenTelemetry is becoming the standard substrate for AI traces, with semantic conventions emerging for prompts, tokens, and tool calls. Building on it keeps your data portable: instrument once against the standard and you are not locked into one vendor’s format, and custom agents, code agents, and web agents can all emit into the same pipeline. It is the lesson distributed tracing taught the microservices era, arriving again for agents.

How Highflame approaches AI observability

Highflame Observatory is the evidence layer over the whole agent fleet: posture score, blast-radius graphs, and compliance evidence in one queryable surface, built on OpenTelemetry and ClickHouse so you query correlated data rather than reassemble logs after an incident. Because every action, an LLM call, an MCP tool call, a file write, already carries the agent’s identity and the authorization decision behind it, Observatory answers the question generic LLM observability cannot: not just what happened, but which agent did it, on whose authority, and whether policy allowed it. See it on the platform page.

Frequently asked questions

What is AI observability?
The practice of capturing and querying everything AI agents do at runtime: LLM calls, MCP tool calls, file reads and writes, and network activity, correlated into sessions and tied to the agent's identity. It goes beyond logging so you can ask why something happened and who did it.
What is the difference between AI observability and LLM observability?
LLM observability is the slice focused on model calls: prompts, completions, tokens, and latency. AI observability is broader: it also covers the tool calls, file changes, and network actions an agent takes, across code agents, web agents, and custom agents. The model call is only one of the things an agent does.
How do you monitor AI agents?
By capturing the agent's behavior, not just its model calls: which tools it invoked with which arguments, which files it touched, the sub-agents it spawned, and the identity it acted under. Tying each action to the agent and its human owner is what turns a stream of activity into an accountable record.
What should you capture from AI agents?
LLM calls (prompts, completions, tokens, latency), MCP and tool calls with arguments and results, file reads and writes, network and web requests, sub-agent delegation, and errors or refusals, each stamped with the agent identity and session that produced it.
What is the difference between monitoring and observability?
Monitoring watches known signals and alerts when one crosses a threshold. Observability lets you ask new questions of correlated data after the fact, like reconstructing a single agent's whole session across every tool and file it touched. Monitoring tells you something is wrong; observability tells you why, and who.

See agent governance against your own agents.