Skip to main content
The Cost Agent is part of managed agents on Enterprise. Through September 15, 2026, Signal and managed agents are included at no additional charge for Enterprise. See Pricing and usage.
The Cost Agent breaks down spend by model, task, and time window, finds expensive models doing low-complexity work, and recommends cheaper routing with quantified savings. Run as an automation, it files ranked cost issues in the same place Signal files its own — scoped to spend rather than correctness. Run as a session, it returns the analysis in the transcript.

​
What it finds

It clusters by cause, so one finding covers a recurring pattern rather than a single expensive call, and each carries the numbers behind it plus a proposed fix.
Cost agent session transcript reporting a 41 dollar total bill across 3,837 LLM spans over 90 days, a cost-by-model table with per-model share and average cost per call, a weekly spend spike, and a top bottleneck recommending a cheaper judge model with the saving quantified

A session run: total spend for the window, a per-model breakdown, and ranked bottlenecks with a costed fix for each.

Every figure comes from your span data; the agent never estimates. It reads an explicit cost attribute where one exists, otherwise multiplies token counts by the model’s published price, and excludes any span with neither — reporting how many it dropped. So accuracy depends on your spans carrying tokens or costs: if token columns show n/a, fix instrumentation first with Track costs.

​
Run it

Signal & Agents → + New Agent → Cost Agent.
Cost Agent setup on the Review and launch step, showing Claude Code as the agent, a bound tracing project, Automation mode on a schedule, and the editable task prompt describing cost breakdown by model and task type

Review & launch: the harness, the project it will read, and the run mode, with the task prompt editable before you create the automation.

You can add context about your cost constraints and approved models; it refines the analysis without overriding the rules above. Most runs reconfirm existing issues and file nothing new, so a run that reports nothing is a valid result rather than a failure.