Tisha Chawla and Susheem Koul argue that model gateways cannot fully control agent costs because runaway spending often emerges from loops, expanding context and sub-agent behavior inside a run. Their proposed control plane attributes every model call to a run and user segment, records cost in a ledger and applies policies at the agent boundary.
Tisha Chawla and Susheem Koul describe lightweight boundary annotations that send inputs, outputs and cost signals to an out-of-band control plane. A governor limits which interventions the platform may apply, allowing policies to compact context, reduce tool output or inject concise instructions before a hard budget cap terminates the run.
Tisha Chawla and Susheem Koul demonstrate preview, halt and steer modes on a two-agent research workflow. Their reported benchmark across open-source agent projects reduced average spend by about 78 percent while improving completion compared with simple throttling, and they propose using accumulated run data to refine or generate policies over time.
Watch on YouTube