Agentic workflows balance task performance against computational cost by using a lower-cost model for execution while selectively consulting a stronger model. In executor-advisor systems the executor determines when to seek advice, but current language models struggle to detect where their own reasoning fails. The need for stronger reasoning emerges during execution and varies across workflow phases.

ARIA: A Meta-Controller for Consultation Decisions

ARIA learns when to consult by balancing task success against consultation cost. The approach formulates consultation as a sequential resource-allocation problem. At workflow contact points ARIA maps execution evidence to a compact state abstraction and selects SILENT or CONSULT, with action values learned by Monte Carlo control from cost-penalized outcomes.

When Stronger Reasoning Is Needed

The need for stronger reasoning emerges during execution and varies across workflow phases. This variability means a static consultation policy cannot optimally trade off performance and cost across diverse tasks and stages of agent operation.

Empirical Results Across Three Benchmarks

Across SWE-bench Verified, Terminal-Bench 2.1, and τ²-bench ARIA recovers 44-81% of the executor-advisor success gap at less than half the advisor model's cost. The policies achieve the best observed success-cost trade-off among ten other evaluated conditions, including routing and cascading methods, fixed and prompted consultation, and random placement at matched advice volume.

Policy Characteristics

The policies are compact and inspectable, with a single consultation-price parameter governing the success-cost trade-off without retraining either underlying model. This design enables rapid adaptation to new tasks or cost structures by adjusting the price parameter alone.

Practical Integration

  1. ARIA sits between the executor model and the advisor model as a decision layer.
  2. At each workflow contact point it observes execution evidence and outputs SILENT or CONSULT.
  3. The consultation-price parameter can be tuned to prioritize either higher success rates or lower computational cost.
  4. No retraining of the executor or advisor is required when the price parameter changes.

Limitations and Future Directions

The framework relies on Monte Carlo estimates, which may require many samples for accurate value learning in sparse-reward settings. The compact state abstraction may overlook fine-grained details that could be relevant in some domains. Future work could explore richer evidence sources, hierarchical consultation policies, or integration with broader agent planning frameworks.

As agentic systems scale to handle more complex and open-ended tasks, the ability to dynamically allocate reasoning resources becomes increasingly critical. ARIA provides a principled, low-overhead approach to this allocation problem, demonstrating that learned consultation policies can capture a substantial portion of the performance gap between cheap and strong models while keeping costs well bounded. Future work may extend the approach to multi-advisor settings, where different advisors specialize on different sub-tasks, or to continuous decision horizons where consultation timing is itself a learned policy. Read the paper on arXiv