ForecastAgentSearch treats world event forecasting as a search problem over specialized AI agents, then openly names four core sub problems any such system would still have to solve.
The hard part of asking AI to forecast a geopolitical event is not the writing. It is knowing which combination of specialized models to trust on a question that mixes regional context, domain expertise, and irreducible uncertainty. A new arXiv preprint called ForecastAgentSearch proposes to treat that selection problem as a search problem, and the more interesting move is that the authors openly publish the four sub-problems they still have not solved.
The framing matters because most AI forecasting pipelines today do not select experts at all. They ask a single large language model to reason through the question, or they stack a chain of models on top of one another and let the strongest final answer win. ForecastAgentSearch, submitted on 30 June 2026 to arXiv's multi-agent systems category, instead starts by analyzing the forecasting query, then searches for and ranks relevant "expert agents" by regional knowledge, domain expertise, reliability, and complementarity, and only then coordinates their analyses into a forecast with explanations and a stated level of uncertainty (ForecastAgentSearch abstract). That ordering is the architectural bet: retrieval over specialized expertise first, coordination second.
The full HTML version of the paper walks through the same sequence in more detail, describing how the system first profiles what kind of geopolitical question is being asked, then looks for agents whose prior training or specialization plausibly matches the region and domain, and only after ranking those candidates does it combine their analyses (ForecastAgentSearch full text). The model is closer to how a research desk might staff a question, asking who knows this region, who knows this sector, and who has been reliable before, than to the standard "ask one model and prompt it harder" pattern.
The contribution that survives past this preprint, even if ForecastAgentSearch itself never ships, is the four open design challenges the paper names in its own abstract: agent profiling, expert retrieval, ranking, and multi-agent coordination. Each is a real engineering question rather than a solved step.
Agent profiling is the question of how you represent what an AI agent actually knows. A region-tagged large language model fine-tuned on Japanese banking regulation is not the same kind of expert as a generic model with a long system prompt that says it is one, and the paper does not yet offer a way to distinguish them.
Expert retrieval is the question of how you find the right agents in the first place. The authors describe a ranking process but do not show how a search step would scale across thousands of candidate agents, or how it would avoid retrieving agents that look relevant but were trained on the wrong slice of the world.
Ranking is the question of how you weigh competing specialist outputs when they disagree. Reliability scoring in the abstract is described as a design goal, not as an evaluated mechanism, so any deployment would need its own calibration story before outputs could be trusted.
Coordination is the question of how multiple agents' analyses are combined into a single forecast. The paper gestures at producing forecasts with explanations and uncertainty awareness, but the actual aggregation mechanism, whether that is weighted averaging, debate, or some other protocol, is left as an open design challenge rather than a specified component (ForecastAgentSearch abstract).
The honest read of the preprint is that it is a framework proposal, not a measured system. The abstract itself describes the work as "an initial step" and lists "possible evaluation protocols for future development" rather than reporting benchmark numbers against human superforecaster aggregates, single-agent LLM baselines, or earlier multi-agent forecasting systems.
A reader evaluating the next wave of multi-agent forecasting claims can lift the four sub-problems off this paper and use them as a checklist. Did the proposed system describe its agent profiling, its retrieval, its ranking, and its coordination, or did it skip one and call the result a forecasting tool? Until a follow-up paper, an institutional deployment, or an independent benchmark actually fills in the four blanks this preprint names, "multi-agent forecasting" is still more of a research agenda than a product category.