Agent cost observability

Provider dashboards and most LLM observability tools are organised around the call or the credential, because that is what the billing system emits. Neither is a unit anyone makes decisions about. Being able to name the most expensive agent, or the run that cost twenty times the median, is what turns a bill into something actionable.

Six agents share one OpenAI key. The provider reports one monthly figure. With agent cost observability the same spend decomposes into six named lines, each with its own cost per run, and the one that regressed last Tuesday is visible as a line that moved.

By attribution coverage — the share of total spend carrying an agent identity rather than landing in an unattributed bucket — and by whether cost can be grouped by run, team, customer, and cost centre without re-instrumenting anything.

What is agent cost observability?

Agent cost observability is the ability to see what each AI agent, run, team, and customer costs, rather than only what an API key costs in aggregate. It requires attribution at the point each call is made, because a shared credential cannot distinguish the agents behind it. The practical test is whether you can name your most expensive agent without reading code.

What is agentic observability?

Agentic observability is observability whose unit is the agent and the run rather than the individual request. Call-level tracing tells you what one request did; agentic observability tells you what one task cost across the forty calls and six agents it triggered. Cost, latency, and failure all need that wider unit in agent systems, because no single call explains the outcome.

How do I get visibility into which agent is costing the most?

Attribute every provider call to a named agent, then rank agents by attributed spend over a period. In Capsera the name comes from a decorator or context manager at the call site, so sub-agents created at runtime inherit their parent's chain automatically. Ranking by cost per run rather than by total is usually more useful, since a cheap agent invoked constantly and an expensive one invoked rarely need different fixes.

How do I get cost visibility into a multi-agent system?

Propagate one run identifier through the whole task and let each agent add its own identity on top, so every call carries both. Cost then rolls up two ways: by run, which tells you what the task cost, and by agent, which tells you which participant drove it. Without the run identifier a multi-agent system reports as a pile of unrelated calls.

• Cost attribution — Cost attribution is the assignment of every unit of LLM spend to the agent, team, feature, or customer that caused it, rather than leaving it aggregated under a shared API key.
• Cost per run — Cost per run is the total LLM spend attributed to one execution of an agent task, summed across every call, model, and agent that execution triggered.
• Agentic spend governance — Agentic spend governance is the practice of attributing, budgeting, and enforcing LLM spend at the level of the agents that generate it, rather than at the level of API keys or accounts.