Language models are increasingly used as agents to help humans decide what information is surfaced. This usage incentivizes content creators to optimize content in ways that appeal not only to humans, but also to agents that mediate access to them. In this paper, we study selection shifts induced by rewriting in agentic decision-making. Given a set of competing content snippets, we rewrite only one snippet while leaving the rest unchanged, and then measure how it impacts the agent’s choice. We operationalize this setup with AgentBait an advisor-rewriter framework in which the advisor learns to propose rewriting strategies and the rewriter revises the snippet. While a rewriter with a fixed prompt improves target snippet selection from 17.1% to 34.8%, our AgentBait raises its selection to 98.5%. We further show that the advisor trained with AgentBait effectively transfers to setups with different agents, languages, and snippets in other domains (e.g., scientific papers). However, higher target selection can reflect unsupported rewrites rather than better content. Adding a reward for support from the original snippet redirects the advisor toward more supported rewriting strategies, revealing a trade-off between factuality and target selection. Together, our results show that once agents mediate access to information, content can be rewritten to be chosen by the agent, even when selection and usefulness diverge.

Vision-language models (VLMs) are increasingly used in multi-image, multi-turn agentic settings where decisions depend on visual changes. However, in existing open-weight VLMs, visual comparisons happen only inside the language model, while the visual encoder itself remains stateless: each image is encoded independently, without access to the prior visual context. As a result, small but task-critical changes may be attenuated before the language model has a chance to compare them, especially when those changes do not affect the high-level semantics of the scene. We introduce a Stateful Visual Encoder, which conditions each visual representation on prior visual features. Under supervised finetuning, VLMs equipped with stateful encoders achieve consistent improvements on controlled tasks involving cross-image spatial aggregation, multi-object visual differencing, and visual trajectory behavior cloning. These improvements are consistent across input resolutions, language model sizes, and VLM backbones. Finally, we validate our model on real-world tasks, including longitudinal radiology, fine-grained image comparison, and remote sensing, where stateful encoders consistently improve generalist VLM baselines and can match or surpass specialized models in selected domains.