Comment l'économie de l'IA multi-agents influence l'automatisation des entreprises

Managing the economics of multi-agent AI now dictates the financial viability of modern business automation workflows. Organisations progressing past standard chat interfaces into multi-agent applications face two primary constraints. The first issue is the thinking tax; complex autonomous agents need to reason at each stage, making the reliance on massive architectures for every subtask too expensive and slow for practical enterprise use. Context explosion acts as the second hurdle; these advanced workflows produce up to 1,500 percent more tokens than standard formats because every interaction demands the resending of full system histories, intermediate reasoning, and tool outputs. Across extended tasks, this token volume drives up expenses and causes goal drift, a scenario where agents diverge from their initial objectives. Evaluating architectures for multi-agent AI To address these governance and efficiency hurdles, hardware and software developers are releasing highly optimised tools aimed directly at enterprise infrastructure. NVIDIA recently introduced Nemotron 3 Super, an open architecture featuring 120 billion parameters (of which 12 billion remain active) that is specifically-engineered to execute complex agentic AI systems. Available immediately, NVIDIA’s framework blends advanced reasoning features to help autonomous agents finish tasks efficiently and accurately for improved business automation. The system relies on a hybrid mixture-of-experts architecture combining three major innovations to deliver up to five times higher throughput and twice the accuracy of the preceding Nemotron Super model. During inference, only 12 billion of the 120 billion parameters are active. Mamba layers provide four times the memory and compute efficiency, while standard transformer layers manage the complex reasoning requirements. A latent technique boosts accuracy by engaging four expert specialists for the cost of one during token generation. The system also anticipates multiple future words at the same time, accelerating inference speeds threefold. Operating on the Blackwell platform, the architecture utilises NVFP4 precision. This setup reduces memory needs and makes inference up to four times faster than FP8 configurations on Hopper systems, all without sacrificing accuracy. Translating automation capability into business outcomes The system offers a one-million-token context window, allowing agents to keep the entire workflow state in memory and directly addressing the risk of goal drift. A software development agent can load an entire codebase into context simultaneously, enabling end-to-end code generation and debugging without requiring document segmentation. Within financial analysis, the system can load thousands of pages of reports into memory, improving efficiency by removing the need to re-reason across lengthy conversations. High-accuracy tool calling ensures autonomous agents reliably navigate massive function libraries, preventing execution errors in high-stakes environments such as autonomous security orchestration within cybersecurity. Industry leaders – including Amdocs, Palantir, Cadence, Dassault Systèmes, and Siemens – are deploying and customising the model to automate workflows across telecom, cybersecurity, semiconductor design, and manufacturing. Software development platforms like CodeRabbit, Factory, and Greptile are integrating it alongside proprietary models to achieve higher accuracy at lower costs. Life sciences firms like Edison Scientific and Lila Sciences will use it to power agents for deep literature search, data science, and molecular understanding. The architecture also powers the AI-Q agent to the top position on DeepResearch Bench and DeepResearch Bench II leaderboards, highlighting its capacity for multistep research across large document sets while maintaining reasoning coherence. Finally, the model claimed the top spot on Artificial Analysis for efficiency and openness, featuring leading accuracy among models of its size. Implementation and infrastructure alignment Built to handle complex subtasks inside multi-agent systems, deployment flexibility remains a priority for leaders driving business automation. NVIDIA released the model with open weights under a permissive license, letting developers deploy and customise it across workstations, data centres, or cloud environments. It is packaged as an NVIDIA NIM microservice to aid this broad deployment from on-premises systems to the cloud. The architecture was trained on synthetic data generated by frontier reasoning models. NVIDIA published the complete methodology, encompassing over 10 trillion tokens of pre- and post-training datasets, 15 training environments for reinforcement learning, and evaluation recipes. Researchers can further fine-tune the model or build their own using the NeMo platform. Any exec planning a digitisation rollout must address context explosion and the thinking tax upfront to prevent goal drift