How to Audit What Your Agent Is Actually Spending

When an agent calls paid endpoints during a workflow run, the purchase decisions happen faster than any human can track. The agent hits a service, receives a result, and moves on. The charge appears in your wallet. By the end of a run, the total is visible but the path is not: you know what you spent, not which call was worth it and which was a mistake.

That is where audit fails. This guide covers what to log, what to watch for, and how to use the aggregate payment data to distinguish normal spending from something worth investigating.

What the aggregate data shows

Before auditing your own agent, knowing what normal spending looks like across the market helps calibrate your expectations.

The paid agent endpoint index measures 226,632 USDC transfers to agent payment addresses in the trailing week, across 1,080 distinct services. Measured by Coinbase telemetry we ingest, not our own probe. The concentration is striking: the top service alone receives 48.8% of all measured transfers. This is not broad-based market activity. Agent payment volume is highly concentrated, with the top service accounting for nearly half of all transfers.

The median endpoint charges $0.01 per call, excluding a small number of implausible outliers. That price feels negligible. At any call volume, the per-call price is rarely the number to watch. The number to watch is the call count.

Of the 37,432 listed paid endpoints, 5,449 (14.6%) have more than one distinct paying buyer in the measured period. Also Coinbase telemetry we ingest. Only 14.6% of listed endpoints have recorded a second distinct customer in the measured period. One implication for auditing: spending on an endpoint outside that 14.6% warrants extra scrutiny, because there is no independent confirmation that the service actually works as described.

Three categories of agent spending

Agent spending falls into three categories, each requiring a different audit approach.

Intentional calls are the ones you wired up explicitly. You chose the endpoint, accepted the price, and built the service call into your workflow. Audit these by verifying the actual price charged per call matches the endpoint's current declared price, and that the service is still the right choice for what your workflow needs. Both can drift independently without any notification to you.

Emergent calls happen when your agent, through tool-use or registry lookup, discovers and calls endpoints you did not wire up explicitly. If your agent can call a registry to find services and then call those services directly during a run, this category is possible. Audit emergent calls by keeping a record of every distinct endpoint your agent has ever called. A new domain appearing in that record without a corresponding deliberate integration decision is a flag worth investigating.

Runaway spend comes from retry loops, unhandled error conditions, or bugs. An agent that retries a failed endpoint call without exit conditions can exhaust a wallet at $0.01 per call before any monitoring catches it. Audit runaway spend by checking for clusters of identical calls within short windows. Retry logic in production workflows should have explicit jitter and hard retry limits; the audit log will surface the cases where it does not.

What to log

A useful audit trail requires five fields per call: timestamp, endpoint domain, amount charged, request ID, and workflow context (which run or task this call was part of). Without request IDs, you cannot trace a charge back to the specific decision that triggered it. Without workflow context, you cannot determine whether a cluster of calls is one tight loop or several different runs converging on the same endpoint.

Log these at the point of payment confirmation, not at the point of receiving the result. If the payment succeeds and the result never arrives (timeout, network failure, service error), you still need the charge in your audit record. Logging only on successful result receipt silently drops charges from failures, which are often the cases worth investigating.

What to watch for

Concentration shift. If your agent begins routing spend toward an endpoint it has never called before, review the routing decision. New services may legitimately outperform old ones, but sudden concentration in an unfamiliar endpoint can also indicate that something in your routing logic changed without your knowledge.

Price drift. If the actual charge per call exceeds the declared price you had recorded for an endpoint, the service has repriced since you set the integration up. Check the current declared price and decide whether the service is still worth calling at the new rate.

Call volume spikes. A sudden increase in call volume without a corresponding increase in workflow runs warrants investigation. Volume anomalies surface earlier than cost anomalies, because the per-call price is stable while the call count is variable.

Single-buyer endpoints in production. If your audit log shows spend on endpoints that have never recorded a second distinct buyer, note them. They may work fine. But they have no independent validation, and they are the category with higher churn risk: more likely to go offline, reprice, or degrade in quality without warning.

The discipline

An audit trail only produces value if you review it. Monthly review is the minimum for agents running production workflows. Set a hard cap at the payment provider level: even if the audit log reveals a problem slowly, the cap limits how far runaway spend can travel between reviews. Hard caps at the provider level cannot be overridden by application code or agent decision-making. Application-level soft limits are useful for visibility but are the wrong place to enforce the floor.

The aggregate data makes one thing clear: agent payment volume is highly concentrated. Your own agent's spending is likely to look concentrated too. What the audit trail tells you is whether that concentration reflects the decisions you made, or decisions your agent made without you.