Prime Agent: A Self-Improving RLM Coding Harness with Persistent REPL and Agent CRUD

Prime Intellect has released Prime Agent, an open-source coding harness designed around two abstractions: the Recursive Language Model (RLM) and the Continual Harness. The goal is to move beyond fixed tool-calling schemas and static sub-agents, letting the agent itself manage its context and evolve its own scaffolding during a session.

Recursive Language Model: Context as a REPL

The RLM treats context as a variable and sub-agent delegation as function calls inside a persistent IPython REPL. This gives the model programmatic access to its own history, sub-agents, and tools—it can write language model programs that act on its context. The REPL stores data in variables, so the agent can work through arbitrarily long sessions without losing access to past information.

Continual Harness: Agent Self-Management

The Continual Harness abstracts the harness's own state—prompts, skills, memory, and sub-agents—as CRUD-able objects. The agent can create, read, update, and delete these during its trajectory. Combined with agent-to-agent messaging, this enables orchestration across sub-agents and even across separate Prime Agent sessions. For example, the agent can spawn persistent sub-agents, message them later, or communicate directly with another Prime Agent instance.

Key Architecture: Persistent Kernel as the Only Tool

Prime Agent uses a persistent IPython kernel as its sole tool. All standard harness features—including sub-agents—are invoked as functions within that kernel. Each sub-agent is itself another Prime Agent instance.

• Background daemon: owns all live sessions over a local socket; you can attach/detach without stopping the agent.
• Agents View: a TUI that recursively lists all root sessions and sub-agents; press ← on an empty prompt to open it.
• Session recovery: each worker process is recoverable; if it crashes, the daemon restores from JSONL and kernel state snapshots.
• Memory management: idle sub-agents unload after 30 minutes, and reload on demand from disk.
• Persistent storage: session history saved as append-only JSONL.

Installation

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

Who It's For

Teams running long-horizon autonomous coding evaluations, researchers doing auto-research, or developers who want a coding assistant that can adapt its own harness.