TokenCast predicts an agent run's remaining token use from composable execution segments and refreshes the estimate as new steps arrive. The representation records each segment's own consumption and the context growth that later calls must reread. Across four task suites and six agent models, the paper reports 14.5% average mean-absolute-error reduction over the strongest comparator. An offline budget-control replay used 21.3% fewer tokens at matched trace completion, with about 32.8 ms of cumulative prediction time per SWE-bench Verified run. The linked repository says code is coming soon and currently has no implementation or license.
TokenCast forecasts agent token cost while a task is still running
2026-09-29Frontier watchWATCH
TokenCast predicts an agent run's remaining token use from composable execution segments and refreshes the estimate as new steps arrive. The representation records each segment's own consumption and the context growth that later calls must reread. Across four task suites and six agent models, the paper reports 14.5% average mean-absolute-error reduction over the strongest comparator. An offline budget-control replay used 21.3% fewer tokens at matched trace completion, with about 32.8 ms of cumulative prediction time per SWE-bench Verified run. The linked repository says code is coming soon and currently has no implementation or license.
What this means for you
For teams enforcing per-task agent budgets, a forecast that updates during execution is more useful than a fixed cap chosen before the run. Watch it and reproduce the forecast in shadow mode once code appears; the savings were measured in offline replay, so it should not control live termination yet.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Method | composable segment cost and context-growth representation |
|---|---|
| Evaluation | 4 task suites, 6 agent models, 96 combinations |
| Error Reduction | 14.5% mean absolute error reduction on average |
| Offline Budget Result | 21.3% fewer tokens at matched trace completion |
| Overhead | 32.8 ms mean cumulative prediction time per SWE-bench Verified run |
| Artifacts | placeholder repository; implementation and license unavailable at review time |
Technical impact 3
Engineering value 4
Adoption signal 1
Maturity 1
Verification cost 4
Risk 2