Coroutine harness cuts agent model calls by 71 percent on CAR-bench

The key novelty isn't just fewer model calls, but the mechanism: agents emit suspended Python programs that resume across evaluator exchanges rather than re-invoking the model for each tool result. This shifts the bottleneck from repeated inference to program state management.

This work sits alongside IterSynth's role-decoupled architecture and agentic-ger's context-aware agent reasoning, both from the same week. Where IterSynth separates planning from synthesis to reduce context noise, this paper decouples inference timing from tool dependencies to reduce latency. Both challenge the sequential ReAct loop that has dominated agent design. The coroutine approach is more granular: it doesn't just restructure reasoning roles, it restructures when the model runs at all. Together these papers suggest the field is moving away from 'call the model, get an action, execute, repeat' toward architectures where agents maintain persistent execution state between model invocations.

If Cerebras or other inference providers publish benchmarks showing coroutine-bridge agents outperform ReAct on real-world tasks beyond CAR-bench within the next six months, that signals production viability. If the median call reduction holds on tasks requiring >10 steps (where sequential dependencies compound), the approach scales; if it degrades, the gains are limited to shallow task trees.