Session boundaries held up better than task tags for agent billing

In firsthand operation of , were tracked for each session. The tracking itself took only an afternoon to build with a small layer of code and a table.

The hard part was deciding where to assign costs when an agent launched , retried work, ran jobs in parallel, or shared across runs. Earlier from practitioners showed that task-level and step-level labels stopped working reliably once were involved.

Defining the session around one billable unit before execution was the approach that continued to hold up. Open questions remain about whether a billing unit can handle retries, who asks for proof of the spending, and whether that person wants only a total or a detailed breakdown.

Key points

  • Per- tracking took an afternoon to build with a small code layer and a table.
  • Task-level and step-level cost labels became unreliable when were added.
  • Linking a session to one billable unit before execution was the most durable approach described.
  • Cost allocation for retries, parallel runs, and shared remains unresolved.
  • The reporting unit should match whether the reviewer wants a total or detailed supporting records.
Read original