Prime Agent: A Self-Improving RLM Agent

Prime Intellect's open-source coding and research agent for general and long-running work, built around two abstractions. The Recursive Language Model (RLM) treats context as a variable and tools like sub-agents as function calls inside a persistent IPython REPL — the model programs its own tool use in Python rather than emitting fixed tool-call schemas. The Continual Harness stores prompts, memories, skill descriptions, and reusable sub-agent specs as durable state the agent can refine through small, evidence-backed edits via /refine.

The repo is a fork of Pi: the packages/coding-agent workspace is literally @earendil-works/pi-coding-agent (v0.7.4), and Prime Agent layers its novelty on top of Pi's agent loop. The monorepo spans four TypeScript workspaces (ai for the provider layer, agent for the loop, tui for the terminal, coding-agent for the CLI) plus prime-agent-runtime, a Python package preloaded into every IPython kernel.

Execution is split across three processes: a client (TUI or headless print/JSON/RPC) owns rendering; a daemon supervisor owns discovery, routing, and cross-agent message delivery; and a session worker owns one root session, its scheduler, and the IPython kernel. Workers and kernels are separate processes for lifecycle and failure containment — not security sandboxes. Sessions persist as append-only JSONL, and the kernel's Python namespace survives resume via a per-variable dill snapshot.

Key surface features: sub-agents spawn with await rlm("task") and return a handle immediately (results arrive via agent_message), sessions keep running in a background daemon and can be reattached, /refine never rewrites the immutable base system prompt, skills are importable Python packages, and automatic compaction plus persistent goals, heartbeats, and bounded autonomous mode keep long tasks moving.