acceptodds
Desk rejected at ICLR 2027

Bandit Limited Discrepancy Search for Agent Harness Configuration

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Coding-agent harnesses decompose work into a task DAG and dispatch each node to a sub-agent, but they fix the per-node configuration—agent type, model, reasoning effort, turn budget—from a plan the language model emits once, and never revisit it. We formalise the resulting problem as agent-DAG configuration selection and introduce BLDS-AF, a discrepancy-limited search over the Hamming ball around that plan that replaces the usual independent-arm bandit with a factored surrogate: a ridge posterior over node-arm and edge-pair indicators, weighted by compute allocation, with a group-wise prior that lets the model collapse onto an additive basis while data is scarce and switch interactions on as evidence accumulates. Every evaluation is pooled across configurations rather than treated as evidence about one configuration alone, and a composed candidate from the surrogate's own argmax is evaluated like any other. BLDS-AF beats the deployed greedy cascade and all fourteen baselines we tried, including Hyperband, TPE, evolutionary search, and UCT; it also beats plan-local coordinate descent by a significant margin, an advantage that grows monotonically with the size of the configuration space and holds at every interaction strength. It beats every plan-blind method on NAS-Bench-Macro with every interval excluding zero, and reaches zero normalised regret at a modest budget on a DAG of live multi-turn coding agents. The gains come from pooling statistical strength across configurations: a pre-flight diagnostic derived from the bias–variance condition of the factored surrogate predicts in advance on which substrates this works and on which it does not.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Multi-Armed Bandits, Limited Discrepancy Search, Agent Harnesses, Agent Configuration, Agentic AI, LLM Agents, Hyperparameter Optimization, Configuration Optimization, Sequential Decision-Making, Search Algorithms, Agent Performance, Automated Agent Tuning, Reinforcement Learning, Black-Box Optimization, Tool-Using Agents, Agent Evaluation, AI Systems Optimization

Related papers

Discussion (0)

Sign in to comment.