Agents Conquer Software, Retail Supply Chains Get Smarter

💻 AI Deployment

Gym-Anything: Turn Any Software into an Agent Environment

The Rundown

Researchers just unveiled Gym-Anything, a framework designed to transform any software application into an interactive environment for AI agents. This notable advance addresses a major bottleneck in developing computer-use agents—the significant time and human effort required to create these environments. Gym-Anything frames environment creation as a multi-agent task itself. A coding agent writes setup scripts, downloads data, and configures software, while an audit agent verifies the setup against a checklist. The team applied this pipeline to 200 software applications, creating CUA-World, a collection of over 10,000 long-horizon tasks. These tasks span domains like medical science, astronomy, engineering, and enterprise systems, all configured with realistic data and train/test splits. CUA-World-Long, a subset of these tasks, features challenges requiring over 500 steps, far exceeding existing benchmarks. Distilling successful trajectories into a 2B vision-language model resulted in performance gains over models twice its size. Furthermore, a separate VLM reviewer improved Gemini-3-Flash's performance on CUA-World-Long from 11.5% to 14.0% by providing feedback on incomplete trajectories.

The details

• Gym-Anything converts any software into an interactive environment for AI agents, reducing manual setup time.
• The framework uses a multi-agent approach for environment creation, with coding and audit agents.
• CUA-World comprises over 10,000 long-horizon tasks across 200 software applications.
• CUA-World-Long includes tasks requiring over 500 steps, significantly longer than existing benchmarks.
• A 2B vision-language model trained on distilled trajectories outperformed larger models.
• An auditing VLM improved Gemini-3-Flash's performance on CUA-World-Long by 2.5 percentage points.

Why it matters

This framework democratizes agent development by lowering the barrier to creating complex software environments. Startups can now explore agent applications across a much wider software landscape, potentially automating tasks in niche industries previously inaccessible due to environment creation costs.

🛒 Retail Automation

Flowr: Scaling Retail Supply Chains with Agentic AI

The Rundown

Flowr, a new agentic AI framework, promises to automate end-to-end retail supply chain operations for large supermarket chains. These operations—spanning demand forecasting, procurement, supplier coordination, and inventory replenishment—are typically high-volume, repetitive, and heavily reliant on manual human effort. Flowr decomposes these complex workflows into specialized AI agents, each with a defined cognitive role, enabling automation of processes that previously demanded continuous human coordination. The framework employs a consortium of fine-tuned, domain-specialized LLMs coordinated by a central reasoning LLM. Crucially, it incorporates a human-in-the-loop orchestration model, allowing supply chain managers to supervise and intervene via a Model Context Protocol (MCP)-enabled interface, ensuring accountability and control. Evaluations show Flowr significantly reduces manual coordination overhead, improves demand-supply alignment, and enables proactive exception handling at a scale unachievable manually. The framework's domain-independent design offers a generalizable blueprint for enterprise-wide supply chain automation.

The details

• Flowr automates end-to-end retail supply chain workflows in large supermarket chains.
• It decomposes operations like demand forecasting, procurement, and inventory replenishment into specialized AI agents.
• The framework uses a consortium of fine-tuned LLMs coordinated by a central reasoning LLM.
• A human-in-the-loop model allows managers to supervise and intervene via an MCP-enabled interface.
• Evaluations demonstrate reduced manual coordination overhead and improved demand-supply alignment.
• Flowr enables proactive exception handling at a scale unachievable through manual processes.

Why it matters

⚖️ Agents

ACE-Bench: Evaluating Agents with Scalable Horizons

The Rundown

ACE-Bench emerges as a new benchmark designed to address critical limitations in existing agent evaluation: high environment interaction overhead and imbalanced task distributions. This benchmark is built around a unified grid-based planning task where agents must fill hidden slots in a partially completed schedule, adhering to both local and global constraints. ACE-Bench offers fine-grained control over two key axes: Scalable Horizons, determined by the number of hidden slots (H), and Controllable Difficulty, governed by a decoy budget (B) that introduces misleading candidates. A key advantage is its Lightweight Environment design, where all tool calls are resolved via static JSON files, eliminating setup overhead and enabling fast, reproducible evaluations suitable for training-time validation. Experiments across 13 models of diverse sizes and families reveal significant cross-model performance variations. The benchmark confirms that H and B reliably control task horizon and difficulty, while also demonstrating strong domain consistency and model discriminability, providing interpretable and controllable evaluation of agent reasoning capabilities.

The details

• ACE-Bench evaluates agents with scalable horizons (H) and controllable difficulty (B).
• It uses a unified grid-based planning task where agents fill hidden schedule slots.
• Lightweight environments resolve tool calls via static JSON files, reducing overhead.
• ACE-Bench enables fast, reproducible evaluations suitable for training-time validation.
• Experiments show H and B reliably control task horizon and difficulty.
• The benchmark exhibits strong domain consistency and model discriminability across 13 tested models.

Why it matters

Reliable and efficient agent evaluation is crucial for rapid development. ACE-Bench's design allows startups to quickly test and iterate on agent performance, especially for long-horizon tasks, ensuring their agents are robust and scalable before costly deployment.

🎨 Generative Graphics

SciTikZer: Advancing Scientific Graphics Synthesis

The Rundown

A new framework tackles the challenge of Scientific Graphics Program Synthesis, aiming to reverse-engineer static visuals into editable TikZ code. While TikZ is standard for scientific schematics, its precision requirements challenge Multimodal Large Language Models. The research addresses a data quality gap with SciTikZ-230K, a large-scale, high-quality dataset covering 11 scientific disciplines, generated by an Execution-Centric Data Engine. An evaluation gap is filled by SciTikZ-Bench, a multifaceted benchmark assessing both visual fidelity and structural logic. To further optimize visual-code generation, a novel Dual Self-Consistency Reinforcement Learning paradigm is introduced, utilizing Round-Trip Verification to penalize degenerate code and boost self-consistency. The trained model, SciTikZer-8B, achieves current best performance, outperforming proprietary models like Gemini-2.5-Pro and large models like Qwen3-VL-235B-A22B-Instruct on this challenging task.

The details

• The framework synthesizes editable TikZ code from scientific graphics.
• SciTikZ-230K is a new dataset with 230,000 image-TikZ pairs across 11 disciplines.
• SciTikZ-Bench evaluates both visual fidelity and structural logic of generated code.
• Dual Self-Consistency Reinforcement Learning optimizes code generation.
• SciTikZer-8B outperforms Gemini-2.5-Pro and Qwen3-VL-235B-A22B-Instruct.
• Round-Trip Verification is used to penalize degenerate code and improve self-consistency.

Why it matters

This advancement in graphics program synthesis can significantly accelerate scientific communication and education. Startups developing tools for researchers, educators, or technical documentation can leverage this to automate the creation of complex diagrams, saving valuable time and ensuring consistency.