Introduction

Run Types shows which kinds of step dominate your agents' runs. It groups spans by span type — for example LLM and TOOL — and shows the top 5.

Where to find it

AI Governance > Monitoring > Run Types.

What this view shows

The product labels the chart "top 5 root-child run types": the spans directly under each run, grouped by span type, for the agent and range you picked. A run that is mostly LLM steps spends its time and money on model calls; a run that is mostly TOOL steps spends it acting on other systems.

Note: Run Types is a step-type breakdown, not a list of kinds of work. It cannot separate two kinds of work done by the same agent — a quick lookup and a long research task both appear as the same mix of LLM and TOOL steps. To compare kinds of request, read individual traces in the agent's Observability tab.

What to do with it

1. See which step type dominates — Mostly LLM steps points to LLM Calls and Cost & Tokens for the next question; mostly TOOL steps points to Tools.

2. Watch for a change in the balance — A run type that suddenly takes a bigger share is a change in how the agent works, worth explaining before it shows up as latency or cost.

3. Narrow to one agent — A fleet-wide breakdown blends very different agents. The agent picker shows the mix for one.

The governance angle

Policy coverage is per asset — a policy cannot be scoped to a step type or to one kind of work. So where two kinds of work genuinely need different rules, the answer is two registry records rather than one agent — the registry record is the unit policy attaches to. See Scoping & Precedence.

Related

• Traces

• Scoping: All Agents vs One Agent

• Sessions, Traces & Spans