Hi FounderCoHo Community,
The foundational abstractions of AI reinforcement learning are fracturing. For years, we treated agent environments as short, stateless, and resettable “episodes”—a legacy inherited from the ATARI era.
Today’s frontier agents don’t live in episodes. Their tasks span days, not seconds. They interact with warm caches, live processes, open sockets, and dirty git trees. Capturing, freezing, and scaling this environment state is the next great engineering hurdle. We are moving past simple prompting into the gritty realities of long-horizon execution and evaluation frameworks that make agents truly trustworthy.
On Tuesday, July 21, FounderCoHo and Daytona are co-hosting another exclusive, high-signal evening at Stanford University. This is a “no-hype” zone engineered for AI researchers, engineers, and founders who are building the infrastructure for the next wave of agentic intelligence.
👉 Request to Join the Event Here: https://luma.com/ai-researchers
Jun Park – CEO at hillclimb
Talk: “We Scaled Data Wrong” Chatbots had Common Crawl; coding agents had GitHub. But the next leap in model intelligence demands hyperspecialized training data (finance, health, law). Jun will dissect where the industry went wrong in data scaling and how we unlock the next tier of specialized data.
Muhammad Annas Hashmi – DevRel at Daytona
Talk: “Today’s Agents Don’t Live In Episodes” Annas will break down how the shift from seconds-long tasks to multi-day sessions stresses our inherited toolkits. He will run a live demo of long-horizon sessionful rollouts, mid-trajectory forking, and cross-calendar-time training using VMs that you can fork cheaply and snapshot mid-run.
Sijun Tan – PhD Researcher at UC Berkeley’s Sky Computing Lab
Talk: “Open RL Stack for Training Long-Horizon Agents” Sijun will introduce rLLM, an open reinforcement learning stack designed to make asynchronous rollouts, custom reward shaping, and multi-step trajectories practical and reproducible without rewriting agent code.
Hanchen Li – PhD Researcher at UC Berkeley’s Sky Computing Lab
Talk: “FrontierCS: Evaluating LLMs on Open-Ended CS Problems” Auto-Research is exploding. As a pioneer in benchmarking this space before it was hot, Hanchen will share the design philosophy behind Frontier-CS and their ongoing work to improve auto-research data pipelines and evaluations.
We are bypassing the high-level fluff. This event is a deep dive into the infrastructure required to transition agents from simple experiments to production-grade autonomy:
The Three Pillars of Scaling: Unpacking rollout horizons (seconds > days), environment states (disposable > first-class learning substrates), and branching (speculative fork trees).
The Hyperspecialized Data Crisis: Identifying the missing infrastructure and methodologies preventing us from acquiring high-quality training data for complex domains like law, finance, and health.
Decoupling Agent Runtimes: How an open stack (rLLM) allows builders to execute agents in flexible sandboxes and plug in custom evaluation logic across diverse training backends.
Benchmarking Auto-Research: Inside the data pipelines and evaluation structures used to stress-test LLMs against highly complex, open-ended computer science problems.
🗓️ Event Details
Date: Tuesday, July 21, 2026
Time: 5:30 PM – 8:30 PM PST (Opening remarks start at 5:30 PM; Catering & Beverages included during the 7:00 PM networking block)
Location: Stanford University, CA (Exact address provided upon approval)
Never miss a FounderCoHo community event, podcast, or newsletter. Join us:
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.