prime-agent is a self-hosted, locally runnable coding and research agent built around a persistent Python control environment and durable harness state. It is designed to continue autonomous work beyond a single chat window.
Project overview
It combines programmatic agent operation, recursive subagents, refinable durable state, and daemon-backed session continuity for work that spans turns and terminal sessions.
Project type
AI Agent · AI Coding · Workflow
Use cases
Coding & Development · Automation
Deployment
Refer to project documentation
License
MIT
Best for
People handling coding or research tasks that need persistent state and continuity beyond a single chat window.
Users who want agent operations, subagent orchestration, and context management exposed through a persistent Python control environment.
Key capabilities
Provides a persistent Python REPL for programmatic file operations, shell commands, tool use, subagent management, and context management.
Spawns child agents for parallel or background work and returns their results programmatically.
Stores supplemental prompts, memories, skill descriptions, and reusable subagent specifications as durable state that can be refined through evidence-backed updates.
Supports importable Python skill packages and includes a skill creator for converting recurring workflows into project or personal skills.
Uses daemon-backed agents that continue running after a terminal disconnect and allow users to reattach later.
Allows running agents to exchange messages and orchestrate one another without routing every interaction through the user.
Uses automatic compaction, persistent goals, heartbeats, schedules, autonomous mode, and retained subagents to preserve progress across turns and terminal sessions.
Limitations and risks
Worker and kernel processes provide lifecycle isolation and recovery, but they do not provide a security sandbox.
Autonomous mode operates within configured turn, token, and time budgets. Reaching a limit does not mean the task succeeded, and a configured verification check confirms only the conditions it evaluates.
The agent executes model-generated Python and project commands with the user's permissions. Untrusted repositories, instructions, skills, or extensions therefore require an external sandbox or restricted environment.
Getting started
Run the provided installer on macOS or Linux, change to the repository or directory where the agent should work, run prime-agent, then run /login and select a subscription or API-key provider.
Evidence and sources
GitHub project description: A self-improving RLM agent for coding workflows and long-running autonomous tasks.
README: Prime Agent is an open-source coding and research agent for general and long-running work.
README: - The **[Continual Harness](https://arxiv.org/abs/2605.09998)** stores supplemental prompts, memories, skill descriptions, and reusable subagent specifications as durable state th…
README: Prime Agent combines a persistent Python control environment with durable harness state, so useful working context and reusable operating patterns can outlive a single chat window.