4,318
Runs metered
Every time an AI agent visits your site, it reads and writes tokens, and someone pays for them. Argus calculates that: token spend and cost per journey, so you can see what each one cost, which ones finished, and which ones spent your money and left with nothing.You cannot govern a channel whose unit cost you cannot state.
Argus is a console, not a dashboard. Every Agentrek simulation and every WARI diagnostic writes a row: what it was, where it ran, whether it finished, the score it produced, and what it cost to produce. Filter by status, by window, by URL or by run id.
Three columns, never one. A run that reads 55,737 tokens and serves 3,751 of them from cache is a different cost profile from one that reads them fresh, and the difference is the only lever you control.
Spend divided by journeys completed, not by runs attempted. The denominator is the value-weighted task that actually finished, which is the same denominator the w0 score uses.
How much of a journey can be served from what the property already published. A low ceiling is usually a documentation problem rather than an infrastructure one.
Wall clock per run, and where inside the journey it accrued. A 666 second run on a 13 call journey is waiting on something, and the ledger says on what.
Every row carries its source: an Agentrek simulation or a WARI diagnostic. Spend rolls up by property, by team key and by run family without a spreadsheet.
Failed runs are priced, cancelled runs are not. Retries are grouped to the attempt that caused them, so a three retry loop reads as one cost, not three unrelated rows.
The expensive runs are rarely the successful ones. A run that walks a journey, pays for identity and discovery, and then meets a dead end has bought tokens and returned no outcome. Argus separates the two so the fix has a price attached.
Two thirds of metered spend is bought by attempts that reached a dead end after paying for discovery, identity and intent.
A journey that cannot complete costs four times more per attempt than one that can, because it retries before it stops.
A card issuer application, 53 calls and 1.66M tokens, abandoned at the confirmation step. Nothing was opened.
Cancellation is free and the ledger says so, which is what makes the priced rows credible.
Argus ships with every Agentrek engagement, and reports against the same frozen instrument the w0 score is measured on.