A multi-agent system (MAS) powered by large language models (LLMs) can automate tedious user tasks such as meeting scheduling that requires inter-agent collaboration. LLMs enable nuanced protocols that account for unstructured private data, user constraints, and preferences. However, this design introduces new risks, including misalignment and attacks by malicious parties that compromise agents or steal user data. In this paper, we propose the Terrarium framework for fine-grained study on safety, privacy, and security in LLM-based MAS. We repurpose the blackboard design, an early approach in multi-agent systems, to create a modular, configurable testbed for multi-agent collaboration. We identify key attack vectors such as misalignment, malicious agents, compromised communication, and data poisoning. We implement three collaborative MAS scenarios with four representative attacks to demonstrate the framework's flexibility. By providing tools to rapidly prototype, evaluate, and iterate on defenses and designs, Terrarium aims to accelerate progress toward trustworthy multi-agent systems.

We present MAFA (Multi-Agent Framework for Annotation), a production-deployed system that transforms enterprise-scale annotation workflows through configurable multi-agent collaboration. Addressing the critical challenge of annotation backlogs in financial services, where millions of customer utterances require accurate categorization, MAFA combines specialized agents with structured reasoning and a judge-based consensus mechanism. Our framework uniquely supports dynamic task adaptation, allowing organizations to define custom annotation types (FAQs, intents, entities, or domain-specific categories) through configuration rather than code changes. Deployed at JP Morgan Chase, MAFA has eliminated a 1 million utterance backlog while achieving, on average, 86% agreement with human annotators, annually saving over 5,000 hours of manual annotation work. The system processes utterances with annotation confidence classifications, which are typically 85% high, 10% medium, and 5% low across all datasets we tested. This enables human annotators to focus exclusively on ambiguous and low-coverage cases. We demonstrate MAFA's effectiveness across multiple datasets and languages, showing consistent improvements over traditional and single-agent annotation baselines: 13.8% higher Top-1 accuracy, 15.1% improvement in Top-5 accuracy, and 16.9% better F1 in our internal intent classification dataset and similar gains on public benchmarks.

Agentic AI systems, which leverage multiple autonomous agents and Large Language Models (LLMs), are increasingly used to address complex, multi-step tasks. The safety, security, and functionality of these systems are critical, especially in high-stakes applications. However, the current ecosystem of inter-agent communication is fragmented, with protocols such as the Model Context Protocol (MCP) for tool access and the Agent-to-Agent (A2A) protocol for coordination being analyzed in isolation. This fragmentation creates a semantic gap that prevents the rigorous analysis of system properties and introduces risks such as architectural misalignment and exploitable coordination issues. To address these challenges, we introduce a modeling framework for agentic AI systems composed of two foundational models. The first, the host agent model, formalizes the top-level entity that interacts with the user, decomposes tasks, and orchestrates their execution by leveraging external agents and tools. The second, the task lifecycle model, details the states and transitions of individual sub-tasks from creation to completion, providing a fine-grained view of task management and error handling. Together, these models provide a unified semantic framework for reasoning about the behavior of multi-AI agent systems. Grounded in this framework, we define 17 properties for the host agent and 14 for the task lifecycle, categorized into liveness, safety, completeness, and fairness. Expressed in temporal logic, these properties enable formal verification of system behavior, detection of coordination edge cases, and prevention of deadlocks and security vulnerabilities. Through this effort, we introduce the first rigorously grounded, domain-agnostic framework for the systematic analysis, design, and deployment of correct, reliable, and robust agentic AI systems.

Current frameworks for consensus statement generation with large language models lack the inherent structure needed to provide provable fairness guarantees when aggregating diverse free-form opinions. We model the task as a multi-objective, token-level Markov Decision Process (MDP), where each objective corresponds to an agent's preference. Token-level rewards for each agent are derived from their policy (e.g., a personalized language model). This approach utilizes the finding that such policies implicitly define optimal Q-functions, providing a principled way to quantify rewards at each generation step without a value function (Rafailov et al., 2024). This MDP formulation creates a formal structure amenable to analysis using principles from social choice theory. We propose two approaches grounded in social choice theory. First, we propose a stochastic generation policy guaranteed to be in the ex-ante core, extending core stability concepts from voting theory to text generation. This policy is derived from an underlying distribution over complete statements that maximizes proportional fairness (Nash Welfare). Second, for generating a single statement, we target the maximization of egalitarian welfare using search algorithms within the MDP framework. Empirically, experiments using language models to instantiate agent policies show that search guided by the egalitarian objective generates consensus statements with improved worst-case agent alignment compared to baseline methods, including the Habermas Machine (Tessler et al., 2024).

Web-based 'deep research' agents aim to solve complex question - answering tasks through long-horizon interactions with online tools. These tasks remain challenging, as the underlying language models are often not optimized for long-horizon reasoning and exploration. Prior work has proposed workflows for constructing instruction-tuning datasets, often leveraging knowledge graphs. However, such methods typically lack fine-grained control over difficulty and quality, yielding synthetic data that falls short of capturing the complexity required for long-horizon reasoning. Furthermore, many studies conflate data and training effects by comparing models trained under different optimization recipes, making it difficult to isolate and evaluate the effectiveness of the data itself. We introduce a two-pronged data synthesis pipeline that generates question - answer pairs by progressively increasing task complexity until a frontier baseline web agent fails. The baseline agent plays multiple roles in this process: attempting the questions, validating factuality, checking for alternative answers, and enforcing filtering. To evaluate the effectiveness of our synthesis methods, we adopt a controlled training setup based on distillation from strong web agents. Experiments across multiple web-based benchmarks show that our dataset - despite being smaller - enables the training of more effective web agents than existing datasets. In particular, our data exhibits twice the diversity in tool-use actions, allowing models trained on it to achieve stronger performance while avoiding repetitive tool-calling behaviors.

Chain-of-thought prompting has popularized step-by-step reasoning in large language models, yet model performance still degrades as problem complexity and context length grow. By decomposing difficult tasks with long contexts into shorter, manageable ones, recent multi-agent paradigms offer a promising near-term solution to this problem. However, the fundamental capacities of such systems are poorly understood. In this work, we propose a theoretical framework to analyze the expressivity of multi-agent systems. We apply our framework to three algorithmic families: state tracking, recall, and $k$-hop reasoning. We derive bounds on (i) the number of agents required to solve the task exactly, (ii) the quantity and structure of inter-agent communication, and (iii) the achievable speedups as problem size and context scale. Our results identify regimes where communication is provably beneficial, delineate tradeoffs between agent count and bandwidth, and expose intrinsic limitations when either resource is constrained. We complement our theoretical analysis with a set of experiments on pretrained LLMs using controlled synthetic benchmarks. Empirical outcomes confirm the tradeoffs between key quantities predicted by our theory. Collectively, our analysis offers principled guidance for designing scalable multi-agent reasoning systems.

Cellular image segmentation is essential for quantitative biology yet remains difficult due to heterogeneous modalities, morphological variability, and limited annotations. We present GenCellAgent, a training-free multi-agent framework that orchestrates specialist segmenters and generalist vision-language models via a planner-executor-evaluator loop (choose tool $\rightarrow$ run $\rightarrow$ quality-check) with long-term memory. The system (i) automatically routes images to the best tool, (ii) adapts on the fly using a few reference images when imaging conditions differ from what a tool expects, (iii) supports text-guided segmentation of organelles not covered by existing models, and (iv) commits expert edits to memory, enabling self-evolution and personalized workflows. Across four cell-segmentation benchmarks, this routing yields a 15.7\% mean accuracy gain over state-of-the-art baselines. On endoplasmic reticulum and mitochondria from new datasets, GenCellAgent improves average IoU by 37.6\% over specialist models. It also segments novel objects such as the Golgi apparatus via iterative text-guided refinement, with light human correction further boosting performance. Together, these capabilities provide a practical path to robust, adaptable cellular image segmentation without retraining, while reducing annotation burden and matching user preferences.

Large Language Models (LLMs) are increasingly used to simulate public opinion and other social phenomena. Most current studies constrain these simulations to multiple-choice or short-answer formats for ease of scoring and comparison, but such closed designs overlook the inherently generative nature of LLMs. In this position paper, we argue that open-endedness, using free-form text that captures topics, viewpoints, and reasoning processes "in" LLMs, is essential for realistic social simulation. Drawing on decades of survey-methodology research and recent advances in NLP, we argue why this open-endedness is valuable in LLM social simulations, showing how it can improve measurement and design, support exploration of unanticipated views, and reduce researcher-imposed directive bias. It also captures expressiveness and individuality, aids in pretesting, and ultimately enhances methodological utility. We call for novel practices and evaluation frameworks that leverage rather than constrain the open-ended generative diversity of LLMs, creating synergies between NLP and social science.

The advent of powerful Large Language Models (LLMs) has ushered in an ``Age of the Agent,'' enabling autonomous systems to tackle complex goals. However, the transition from prototype to production is hindered by a pervasive ``crisis of craft,'' resulting in agents that are brittle, unpredictable, and ultimately untrustworthy in mission-critical applications. This paper argues this crisis stems from a fundamental paradigm mismatch -- attempting to command inherently probabilistic processors with the deterministic mental models of traditional software engineering. To solve this crisis, we introduce a governance-first paradigm for principled agent engineering, embodied in a formal architecture we call ArbiterOS.

Real-world data analysis tasks often come with under-specified goals and unclean data. User interaction is necessary to understand and disambiguate a user's intent, and hence, essential to solving these complex tasks. Existing benchmarks for evaluating LLMs on data analysis tasks do not capture these complexities or provide first-class support for interactivity. We introduce ConDABench, a framework for generating conversational data analysis (ConDA) benchmarks and evaluating external tools on the generated benchmarks. \bench consists of (a) a multi-agent workflow for generating realistic benchmarks from articles describing insights gained from public datasets, (b) 1,420 ConDA problems generated using this workflow, and (c) an evaluation harness that, for the first time, makes it possible to systematically evaluate conversational data analysis tools on the generated ConDA problems. Evaluation of state-of-the-art LLMs on the benchmarks reveals that while the new generation of models are better at solving more instances, they are not necessarily better at solving tasks that require sustained, long-form engagement. ConDABench is an avenue for model builders to measure progress towards truly collaborative models that can complete complex interactive tasks.