​From Minutes to Days: The Rise of Long-Horizon AI

​Gad Benram
Co-Founder and CTO, TensorOps

​Abstract

​AI is moving from answering questions to doing work. What began with prompts and short, isolated tasks is becoming something fundamentally different: agents that can operate for hours or days, coordinate across multiple steps, use tools, recover from failures, and continue working toward a goal with limited human involvement.

​This opening talk will explore the transition to long-horizon AI, what is making it possible, and what changes when an agent is expected to keep working rather than simply produce an answer. Gad will cover the techniques behind this shift, including harness engineering, tool use, state and memory, verification, feedback loops, and multi-agent collaboration, along with the use cases that become possible as the time horizon expands.

​He will also examine recent examples that show how quickly these capabilities are advancing. During an OpenAI cybersecurity evaluation, autonomous agents escaped their test environment and compromised Hugging Face infrastructure, sustaining a complex, multi-step operation over several days. The incident offers a striking example of agents carrying out work that would traditionally require extended human effort.

​The talk will set the stage for the series by asking a broader question: What happens when AI systems can keep working long after the prompt is over?

​About Gad

​Gad Benram is Co-Founder and CTO of TensorOps. He works at the intersection of AI engineering and production infrastructure, with a focus on the systems, tools, and architectures required to make increasingly autonomous AI agents useful in the real world.

​Agents don't just forget facts. They forget WHY.

​Ori Siegal
Founder

​Abstract

​Long-horizon agents do not usually fail because the model suddenly gets worse at step 400. They fail because step 400 can no longer see why step 12 made a particular decision.

​Out-of-the-box memory systems help preserve information, but often lose the reasoning that gave that information meaning: what was debated, what was accepted or rejected, what assumptions were made, and what was ultimately concluded. The result is an agent working from a plan that has outlived its reasoning, drifting while still appearing healthy.

​This talk explores memory as a governance problem rather than simply a storage problem. Ori will cover what knowledge is worth preserving and in what form, what should be discarded early, how to determine whether information is still valid, and what makes a rule or piece of knowledge actually influence an agent when it matters.

​The methodology comes from practice. Ori has spent the past four months building and operating a system like this every day, and will share what held up, what broke, and what he would do differently.

​About Ori

​Ori Siegal has spent more than a decade working with data, from data science to agentic systems and the infrastructure they run on. Today, he focuses on memory and knowledge management for AI agents.