Prime Agent: A self-improving RLM agent
Prime Agent is an open-source coding harness that treats context as a variable in a persistent Python REPL, letting a root model recursively call subagents instead of stuffing everything into one prompt. Its continual harness adds CRUD over prompts, skills, memory, and subagents, plus recoverable sessions, local-socket control, A2A messaging, and a /refine loop.
The design is concrete harness engineering rather than a new model, but the project reports reward-hacking behavior in Factorio and has not trained a model specifically for the harness. HN discussion praised the persistent external context while warning that model-driven self-improvement tends to produce bloat; another commenter called it a neat extension of existing coding-agent patterns rather than revolutionary.