MARGENT: Measuring the Marginal Value of Delegation in Agentic Systems

Agentic systems increasingly rely on specialized agents to improve reasoning, but each additional interaction consumes computation and may leave an answer unchanged or even corrupt a correct one. Existing systems often follow fixed interaction patterns or select calls using query features and model confidence. These signals do not directly establish whether another agent can improve the candidate answer already under consideration. This raises a fundamental question: When is additional intelligence worth it? We introduce MARGENT, a trainable framework for a manager that must either commit to its current answer or delegate to a sub-agent. By comparing commitment and delegation from the same manager-generated state, MARGENT measures the marginal value of a call and distills shortest successful branches into supervision for useful delegation and timely stopping. Re-collecting these branches after manager updates extends the same principle to recursive self-improvement. Across four reasoning benchmarks, the best single-call oracle captures 91.7% of the measured multi-call oracle gain on average. Trained managers improve their own pre-delegation accuracy by 17.1 percentage points with 0.53 sub-agent calls per question. On development data, roughly one selected call matches three forced calls on MMLU-Pro and outperforms them on AQuA-RAT. Iterative experiments further examine how refreshing supervision adapts delegation to a changing manager. These results support allocating additional computation according to its marginal value at the current reasoning state, with both useful calls and timely commitment learned from measured outcomes.