wtf is prime-agent?

Prime Agent is an open source coding and research agent built for work that runs for a long time, not just a single quick chat exchange. Instead of a normal one-shot conversation, it runs inside a persistent Python environment, so the code, variables, and state from earlier in a session stay available later, and it can spawn child agents to handle pieces of work in parallel or in the background and then report their results back. A distinctive part of the design is that Prime Agent can improve its own working notes over time. A command called refine reviews what happened in a session and can make small, evidence backed updates to its stored prompts, memories, and reusable subagent instructions, while never touching its core underlying instructions, and every change is recorded so it can be rolled back. Reusable skills are just importable Python packages, and there is a built in tool that can turn a workflow you repeat often into a saved skill for later use. Sessions run through a background service, called a daemon, so a running agent keeps working even if you disconnect your terminal, and you can reattach to it later from the command line. Prime Agent also supports persistent goals that stay active across many turns until finished, scheduled or heartbeat style check-ins that bring it back to a task periodically, and a bounded autonomous mode that runs within limits you set for turns, tokens, and time, optionally checked against quality gates you define yourself. Installation is a single shell script for macOS or Linux that downloads a verified release and installs the prime-agent command. On first run you log in with either a subscription or an API key, then start the agent from inside the project directory you want it to work on. The README is explicit that Prime Agent runs model generated code and commands with your own user permissions, so it is not a security sandbox, and recommends working in a disposable clone or checkpoint you can inspect and restore, and only using trusted repositories, instructions, and skills. Written primarily in TypeScript, the project is released under the MIT license.