agentsfactbullishA bounded contract approach using typed retrieval with no raw cross-decision transcript improves long-horizon LLM agent performanceComputation and Language27 Jul 2026http://arxiv.org/abs/2607.02255v1