THE INFERENCE HUB - Long Horizon AI and Context Extension in LLMs

AI is moving from answering questions to 𝗱𝗼𝗶𝗻𝗴 𝘄𝗼𝗿𝗸. In this webinar from 𝗧𝗵𝗲 𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗛𝘂𝗯, Gad Benram and Ori Siegal explore what happens as AI agents move from short, isolated tasks to long-horizon systems that can operate for hours or days — using tools, maintaining context, recovering from failures, and continuing toward a goal with limited human involvement. 𝗙𝗥𝗢𝗠 𝗠𝗜𝗡𝗨𝗧𝗘𝗦 𝗧𝗢 𝗗𝗔𝗬𝗦: 𝗧𝗛𝗘 𝗥𝗜𝗦𝗘 𝗢𝗙 𝗟𝗢𝗡𝗚-𝗛𝗢𝗥𝗜𝗭𝗢𝗡 𝗔𝗜 𝗚𝗮𝗱 𝗕𝗲𝗻𝗿𝗮𝗺 — Co-Founder & CTO, TensorOps What does it take to build agents that keep working long after the initial prompt? Gad explores the techniques enabling long-horizon AI, including harness engineering, tool use, state and memory, verification, feedback loops, and multi-agent collaboration. He also looks at recent examples of autonomous agents sustaining complex, multi-step operations over extended periods — and what these advances mean for the future of AI engineering. 𝗧𝗵𝗲 𝗯𝗶𝗴𝗴𝗲𝗿 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻: What happens when AI systems can keep working long after the prompt is over? 𝗔𝗚𝗘𝗡𝗧𝗦 𝗗𝗢𝗡’𝗧 𝗝𝗨𝗦𝗧 𝗙𝗢𝗥𝗚𝗘𝗧 𝗙𝗔𝗖𝗧𝗦. 𝗧𝗛𝗘𝗬 𝗙𝗢𝗥𝗚𝗘𝗧 𝗪𝗛𝗬. 𝗢𝗿𝗶 𝗦𝗶𝗲𝗴𝗮𝗹 — Founder Long-horizon agents don’t necessarily fail because the model gets worse after hundreds of steps. They fail because the agent can no longer see 𝘄𝗵𝘆 earlier decisions were made. Ori explores memory as a governance problem rather than simply a storage problem: what knowledge should be preserved, what should be discarded, how to determine whether information is still valid, and how to make sure important rules and decisions influence an agent when they matter. Drawing on months of hands-on experience building and operating these systems, Ori shares what worked, what broke, and what he would do differently. 🔗 𝗧𝗛𝗘 𝗜𝗡𝗙𝗘𝗥𝗘𝗡𝗖𝗘 𝗛𝗨𝗕 A community for practitioners shipping AI in production. Website: https://theinferencehub.com/ Upcoming events: https://luma.com/inference ⚙️ 𝗛𝗢𝗦𝗧𝗘𝗗 𝗕𝗬 𝗧𝗘𝗡𝗦𝗢𝗥𝗢𝗣𝗦 AI engineering from strategy to production. https://tensorops.ai/ #AI #AIAgents #AgenticAI #AIEngineering #LongHorizonAI #TheInferenceHub #TensorOps