Prime Agent – an open‑source self‑improving AI coding assistant

What it is

• Prime Agent is a command‑line tool that lets a large language model act as a coding and research agent.
• It builds on two ideas from the authors’ research:
  1. Recursive Language Model (RLM) – the model treats the prompt as a mutable set of variables and can call sub‑agents (other model instances) as if they were functions.
  2. Continual Harness – a durable store for supplemental prompts, memories, skill definitions and sub‑agent specs that can be refined over time.

Key capabilities

• Persistent Python REPL – the agent runs inside a long‑living Python interpreter, so code, file operations, shell commands, and tool calls are executed directly.
• Sub‑agents – rlm.spawn(...) creates child agents that work in parallel or in the background and return results programmatically.
• Skill system – reusable workflows can be packaged as Python packages; the built‑in skill creator turns a pattern into a “skill” that the agent can import later.
• Background daemon – sessions keep running after you detach the terminal; you can re‑attach with prime-agent attach <id>.
• Agent‑to‑agent messaging – running agents can discover each other and exchange messages without user mediation.
• Continual refinement – the /refine command lets the harness store small, evidence‑backed updates (new prompts, memories, skill specs) while keeping the immutable base system prompt untouched. Snapshots enable rollback.
• Autonomous mode – /autonomous lets the agent keep working within configurable turn, token and time budgets, optionally gated by user‑defined quality checks.
• Scheduling & heartbeats – built‑in commands (/heartbeat, prime-agent schedule) let the agent re‑enter a session on a timer or at a specific time.

Typical workflow

1. Install with the one‑line script (Linux/macOS) or PowerShell installer (Windows).
2. Run prime-agent inside a project directory.
3. Log in (/login) to provide an API‑key or subscription for the underlying LLM.
4. Issue commands such as rlm.spawn(...), /goal, /refine, or let the agent run autonomously.
5. When you disconnect, the daemon keeps the Python state, files, and sub‑agents alive; later you can prime-agent attach to continue.

Safety notes

• The tool executes model‑generated Python with the same permissions as the user; it is not sandboxed. Users are advised to run it only on trusted codebases or inside an external sandbox.

Installation

curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh
# Windows
irm https://app.primeintellect.ai/prime-agent/install.ps1 | iex

Main commands

• prime-agent agents – list running/idle/saved sessions
• prime-agent attach <agent> – re‑attach to a live session
• prime-agent --resume [path|id] – resume a specific session
• prime-agent status – view daemon health
• prime-agent doctor [--fix] – diagnose/repair services
• prime-agent update – pull the latest release
• prime-agent shutdown – stop all agents and background services

Where to learn more

• Blog post on the RLM concept: https://www.primeintellect.ai/blog/rlm
• The research paper describing the Continual Harness: https://arxiv.org/abs/2605.09998
• The accompanying arXiv pre‑print for Prime Agent itself: https://arxiv.org/abs/2608.23552

License & citation

• MIT License.
• Cite the work with the provided BibTeX entry if used in research.

---

Prime Agent is a genuine open‑source project aimed at enabling long‑running, self‑improving AI‑driven coding and research workflows.