Agents Under Pressure: Failure Modes, Cost Tradeoffs, and the Race to Production

AgentCost: Benchmarking the Token Efficiency of Autonomous Agent Workflows

New paper measures token spend across popular agent frameworks and finds 3-10x cost variance for equivalent task completion. The efficiency gap between naive and optimized orchestration is real and measurable.

Hierarchical Planning Improves Long-Horizon Agent Task Completion by 40%

Researchers show that breaking long tasks into explicit hierarchical subplans before execution significantly outperforms flat prompting approaches on complex benchmarks. Practical architecture insight with clear implementation implications.