Prime Agent Puts the Coding Agent Inside a Python REPL

Prime Agent topped GitHub's daily trending list, and the reason is worth more than the star count. It's an open-source coding agent built on an idea most agent frameworks avoid: instead of orchestrating tool calls from the outside, put the model inside a persistent Python session and let it write code to do everything.

Two ideas doing the work

The first is the Recursive Language Model. Context is held as variables rather than stuffed into a prompt, and recursive subagents are invoked as function calls, all inside a persistent REPL. If your agent can hold a large result in a variable and call a subagent on part of it, the context window stops being the binding constraint on task size.

The second is the Continual Harness. The agent keeps supplemental prompts, memories and reusable patterns, and refines them through evidence-backed updates. That's the self-improving part — not the weights changing, but the scaffolding around them getting better at this particular repository.

What's in the box

• A persistent IPython environment, so state survives between steps
• Built-in subagents for parallel or background work
• Skills as importable Python packages, rather than prose instructions
• A background daemon, so sessions can detach and reattach
• Direct agent-to-agent communication
• An autonomous mode with configurable budgets and quality gates

The agent runs in your project directory and executes model-generated Python and shell commands. Sessions support scheduled heartbeats and persistent goals, and can spawn child agents.

Install is a shell script from app.primeintellect.ai, then run prime-agent in your target directory and /login. MIT licensed, around 11.4k stars.

The honest assessment

Executing model-generated code in a persistent environment with configurable budgets is powerful and is also exactly the shape of thing that needs a sandbox and an egress policy before it goes anywhere near a corporate repository. "Autonomous mode with budgets" controls spend. It does not control reach.

That caveat aside, the architectural bet here is a good one. Most agent frameworks are elaborate machinery for deciding which tool to call next. Prime Agent's answer is that the model already knows a general-purpose way to call things — it's called Python — and the framework's job is to make that environment persistent and to remember what worked.

Source: PrimeIntellect-ai/prime-agent on GitHub