Datasets:
Pre-registered cost study on AgentLogs (424k sessions): results and reproducible code
Hi, thanks for publishing AgentLogs. We used all 276 log shards (revision 04013a44) for a pre-registered study of whether the cost of a coding-agent run can be predicted before the run starts.
Short version: model x trigger x prompt-length groups narrowed the 80% cost range in 5 of 15 groups on a locked test. Prompt text, repository language and size, and the user's own history added nothing reliable. A per-group p10 lower bound was calibrated (met by 90.0% of locked-test sessions).
One thing that may matter to other users: 31,110 sessions use models without a public list price, so dollar figures exclude them.
If our parsing misreads any field, we'd like to know and will correct it publicly.
Article: https://agentrecourse.org/journal/agent-run-cost
Code and pre-registration: https://github.com/vhsgreed/cost-predictor-spike
Karl Sundström
Update (2026-10-10): a third pre-registered round on the same split added a learned during-run forecast. A LightGBM quantile model beat the turns-only table on point error in all three benchmark sets we use (30%, 18%, 8% lower), but its 80% intervals covered only 75%, 62% and 76%, so it failed our bar on calibration. A stop rule from a success classifier cut estimated tokens per resolved task by 29% on one set and did little on the other two. Numbers and the pre-registration are in agentlogs/PLAN.md round 8 and agentlogs/forecast_r8.out in the same repo.