See agents as agents.
Understand runs, prompts, models, tools, policies, tokens, latency and behavior through an observability layer built for agentic systems.
Infrastructure monitoring wasn't built for agents.
Traditional monitoring tells you whether the infrastructure is healthy. Operating agents means understanding what the agent actually did: which path it took, which tool it called, which rule was evaluated and what it cost. A healthy server can be running an agent that gets things wrong.
Here the trace is not added instrumentation: every agent step already is the span, the event and the checkpoint. One step feeds the trace, the cost ledger, the trail and the memory.
Track the signals that describe agent behavior.
Agents
Which agent ran, on which version and under which owner.
Runs
Every run from start to outcome, with the path it took.
Tasks
The units of work inside a run, including subagents.
Prompts
The exact prompt that was sent, not the template that produced it.
Models
Which model responded, with which parameters and at what cost.
Tools
Which calls went out, with which arguments, and what came back.
Policies
Which rules were evaluated and which one blocked.
Tokens
Consumption per run, agent, team and environment.
Latency
Where the time went, step by step, not just in total.
Outcomes
The outcome of the run, and whether it delivered what was asked.
Waits
Every pending wait, for a document, a callback or an approval, visible in the console. The wait nobody sees is the 3 a.m. incident.
A fleet, not a pilot.
Runs, cost and latency together
Active runs, success rate, p99 and token spend per agent and per tenant. Telemetry tied to money, not just logs.
Fleet health per agent
A sick agent shows up in its error rate and cost before it becomes an incident, with incidents and budget alerts next to the metrics.
Approvals are an operational signal
Pending four-eyes requests are counted alongside the metrics. Governing is part of running the fleet.
The run that ended well and cost ten times more.
01
The agent returned the right answer, and as far as traditional monitoring is concerned, all is well.
02
Along the way it ran the same search eleven times.
03
And escalated to an expensive model for a trivial task.
04
Seeing agents as agents means treating the run as the unit. That's where 'why was this so expensive?' gets an answer.