PrimeIntellect-ai/prime-agent Usage Examples: CLI, RLM, and Custom Skills

PrimeAgent is a self-improving, long-running AI coding assistant that you launch via prime-agent.sh, control through CLI commands like agents, attach, and /refine, and extend using Python skills and the rlm() sub-agent runtime.

PrimeAgent implements a Recursive Language Model (RLM) architecture that treats prompts as mutable variables and enables programmatic sub-agent calls, alongside a Continual Harness for persistent memory and skill storage. This open-source toolkit from PrimeIntellect-ai allows developers to run durable coding sessions that survive disconnects and evolve through iterative refinement.

Installation and First Launch

Getting started requires downloading the entry-point script and launching the daemon-backed service.

Install the latest stable release for macOS or Linux:

curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh

Start the agent from any project directory:

cd /path/to/project
prime-agent

The prime-agent command invokes prime-agent.sh, which boots the background daemon, attaches the IPython kernel, and initializes the TUI client according to the source in /cache/repos/github.com/PrimeIntellect-ai/prime-agent/main/prime-agent.sh.

Core CLI Commands

PrimeAgent provides a comprehensive command-line interface for session management. These commands are dispatched through packages/ai/src/cli.ts.

List, attach, and manage persistent sessions:

prime-agent agents                   # List running, idle, and saved sessions

prime-agent attach <agent>           # Re-attach to a running session

prime-agent --resume [path|id]       # Browse or resume a session

prime-agent status                   # Inspect background service state

prime-agent doctor [--fix]           # Diagnose / repair services

prime-agent update [--force]         # Update the binary

prime-agent shutdown [--force]       # Stop all agents & services

These commands interface with the Coding-Agent layer, which manages the RLM runtime and session persistence.

Interactive Session Workflows

Once inside the TUI, you interact with the Continual Harness—a durable store for supplemental prompts, memories, and skill definitions located in .prime-agent/ at runtime.

Authenticate with your LLM provider:

/login

Refine the harness without touching the immutable base system prompt:

/refine

Running /refine opens a guided interface to edit memories or skill specifications, storing refinements in JSON files under .prime-agent/ as implemented in the harness logic referenced in the README.

Programmatic Usage with RLM

The Recursive Language Model (RLM) enables programmatic sub-agent calls within the persistent REPL. This architecture, detailed in the README overview, allows the model to spawn background workers via function calls.

Spawn a child agent to run isolated tasks:

# Spawn a child agent that runs a background task

result = rlm("run_long_task.py", timeout=300)
print(result)   # The sub-agent returns its final output when finished

The rlm() helper creates a dedicated worker process for the given script and returns results programmatically, enabling complex multi-step workflows where sub-agents operate as functions.

Creating Custom Skills

Extend PrimeAgent by authoring reusable skills—installable Python packages that expose specific capabilities.

Create and package a new skill:

# In a new directory, create a Python package named myskill

mkdir myskill && cd myskill
python -m pip install --quiet --upgrade build

# Write a module that implements a function `run`

# Example: myskill/__init__.py

def run(task):
    # custom logic here

    return f"Completed: {task}"

# Build and install the skill

python -m build
pip install .

Invoke the skill from within PrimeAgent:

/run myskill.run "process dataset.csv"

Skill definitions are stored in the persisted harness and can be refined using /refine to update their behavior across sessions.

Key Source Files and Architecture

Understanding the repository structure helps when customizing or debugging PrimeAgent usage patterns.

Component	Responsibility	Key Source
Entry Point	Boots daemon and UI	prime-agent.sh
CLI Parser	Command dispatch and argument handling	packages/ai/src/cli.ts
LLM Streaming	Provider normalization (OpenAI, Anthropic, Bedrock)	packages/ai/src/stream.ts
Providers	Vendor-specific response handlers	packages/ai/src/providers/*
TUI	Terminal interface and rendering	packages/tui/src/*
Documentation	Usage guides and skill authoring	packages/coding-agent/docs/usage.md, skills.md

The AI layer in packages/ai/src/stream.ts normalizes streaming responses across different LLM providers, ensuring consistent behavior regardless of which /login provider you configure.

Summary

• Launch: Use curl to install and run prime-agent from any project directory to start the daemon and TUI.
• Manage Sessions: Control long-running agents via prime-agent agents, attach, and --resume commands.
• Persist State: Store memories and prompts in the Continual Harness using /refine, saved to .prime-agent/ JSON files.
• Spawn Sub-Agents: Call rlm() programmatically to create isolated worker processes that return structured results.
• Extend Functionality: Author Python skills with a run() function, install them as packages, and invoke via /run.

Frequently Asked Questions

How do I resume a PrimeAgent session after disconnecting?

Use prime-agent --resume to browse available sessions or prime-agent attach <agent> to reconnect to a specific running session. The Continual Harness persists all state to .prime-agent/ files, allowing you to pick up exactly where you left off even if the terminal closes.

What is the difference between the RLM and a standard LLM call?

The Recursive Language Model (RLM) treats the prompt as a mutable variable and allows the model to programmatically invoke sub-agents via rlm() function calls. Unlike standard one-shot LLM calls, RLM creates persistent worker processes that can run long tasks, return structured data, and maintain state across multiple turns within the same session.

Where are custom skills stored in PrimeAgent?

Custom skills are stored within the Persisted Harness under the .prime-agent/ directory created at runtime. Skill definitions, refinement history, and supplemental prompts are saved as JSON files, ensuring they survive between sessions and can be updated via the /refine command without modifying the base system prompt.

Can I use PrimeAgent with cloud providers like AWS Bedrock?

Yes. The AI layer in packages/ai/src/stream.ts and the provider implementations in packages/ai/src/providers/* normalize streaming responses across OpenAI, Anthropic, AWS Bedrock, and other LLM services. Authenticate using /login and the system handles provider-specific formatting automatically.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}