PrimeIntellect-ai releases open-source self-improving RLM coding agent: Prime Agent

Prime Agent is built around two abstractions: the recursive language model (RLM), which treats context as variables and subagent calls as function calls inside a persistent REPL; and the Continual Harness, which stores supplemental prompts, memory, skill descriptions, and reusable subagent specs as persistent state, updated in small evidence-driven steps via /refine without rewriting the immutable base system prompt, with snapshot rollback. Everything is programmatic—files, shell, tools, and subagents all run through code, and rlm.spawn(...) can launch subagents in parallel or in the background. For research teams running long tasks, it offers daemon-backed session resumption: after the terminal disconnects, the agent keeps running and can be re-attached, with heartbeats, scheduled runs, persistent goals, and a bounded autonomy mode to maintain progress. Auto-compaction defaults to a 95% threshold, adjustable in 5-point steps in /settings, or per-model via models.json modelOverrides.<id>.contextWindow, which sets the effective context window for a single model. SSH and tmux users should note the clipboard path: in tmux it relies on the client declaring Ms capability to forward the buffer, and an outer tmux set to set-clipboard external may reject the inner OSC 52—switch the outer to set-clipboard on. macOS Terminal.app does not support OSC 52. Enterprise intranet users can point NODE_EXTRA_CA_CERTS at a PEM bundle of custom CAs, but must run prime-agent shutdown to restart the background daemon for it to take effect. The project is MIT-licensed, and the install script verifies checksums for Windows x64 builds.