ReForm

Reform is a self-building, self-healing agent platform that preserves user intent. Users describe a task, review its goal, success checks and permissions, then build, test and deploy their agent. The platform is designed to support custom agents and teams created through plain language. A hybrid Supervisor coordinates Builder, Tester and Healer agents. It combines predictable workflow controls with model reasoning to investigate failures, propose repairs and independently validate new versions. Repairs cannot silently change the goal, weaken checks or expand permissions; unresolved failures require human review. Reform integrates GLM 5.3 and NVIDIA Nemotron through Crusoe’s managed inference, NVIDIA OpenShell for isolated microVM execution, and scoped local repair memory. Our prototype completed one verified live workflow using GLM 5.3: build, deploy, run, encounter a controlled integration failure, repair the agent and resume the same task under unchanged success checks. Integrations and demonstration data are fictional; general agent-team creation remains part of the product vision. Reform targets Done Means Done, with Version N+1 through independently validated agent repairs.