Stateful environments for vertical agents
summary
Repository implements a stateful environment layer that gives games and apps a same interface for Pokemon Red Minecraft Swe-Bench, isolates logic into a network-isolated environment, and supports reset and rollback so agents can branch, run tree-search over hundreds of steps, compare branches (one Minecraft branch outperforms the other in the demo), and converge on the best result. The code runs as a separate process, exposes compact observations and actions, enables multi-agent setups, and makes swapping models simple for long-horizon vertical workflows.
AI-written summary · about
related talks
26mHow to Build and Govern Production-Grade Multi-Agent Systems
01h 09mMinecraft as AI Playground and Laboratory
14mMake Agents Remember the Right Context at the Right Time
39mAgents for the Sake of Happiness
26mBeyond Sandboxes: Architecting Durable Runtimes for AI Agents
44mThe Swarm Next Door: Running Parallel AI Agents on a Real Android Codeba…
46mVibecoding your Own Multi-Agent Workstation
25mDirecting a Swarm of Agents for Fun and Profit