A research team has released OpenTinker, an open infrastructure for training multiple AI agents on a shared base model without their individual learning states colliding.
Modern agent training pipelines are messy. They blend supervised fine-tuning, reinforcement learning, rollout generation, and multi-turn environment interaction — often all at once. OpenTinker addresses that mess by treating LoRA adapters not as static inference artifacts but as live policy states that carry optimizer data, rollout snapshots, and training attribution. The system lets users define environments, agents, and learning objectives while it handles checkpoints, policy-version refreshes, and rollout sampling under the hood. A unified data path supports both supervised and RL training by converting trajectories into token sequences with explicit masks.
The practical upshot is multi-tenant agent training: many users, tasks, or agents can share a single base model while keeping their adapter updates and rollout snapshots isolated from one another. That matters because the compute cost of running separate base models per agent is prohibitive at scale, and the existing alternatives tend to let training runs interfere with each other in subtle ways.
OpenTinker does not claim to have solved agentic training — it claims to have organized the problem better. Whether that organizational clarity survives contact with production workloads is the question the research paper leaves open.