Stateful environments for vertical agents — Josh Purtell, Synth Labs

Summary

• Josh Purtell (Synth Labs founder) defines stateful environments as engines that compute results external to the agent, exposing only what the agent needs to observe/manipulate — building on the lineage RL Glue → OpenAI Gym → SWE-bench/SWE-agent's 'agent-computer interface'.
• Argues vertical agents (finance, accounting, health) need stateful environments now because models like Sonnet 3.5 enable long-horizon iteration on artifacts (post-Claude-artifacts era), unlike early LLMs that just calculated sums.
• Key benefit: containerizing logic into stable code means you can swap to a new model without rewriting agent logic; supports multi-agent setups via network boundaries (same trick as async-work-in-production).
• Resetable state enables rollbacks and effectively gives you Language Agent Tree Search (LATS, the 2023 paper) for free — agent branches in Minecraft and the best branch wins, avoiding derailment over hundreds of steps.
• Open-sourced abstractions on GitHub at synth-ai/environments with implementations across academic benchmarks.