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 | shWho 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.
📖 Read the full source: HN AI Agents
👀 See Also
Modulus: Cross-repository knowledge orchestration for AI coding agents
Modulus is a desktop app that runs multiple AI coding agents with shared project memory across repositories. It solves cross-repo context problems by letting agents understand dependencies between different codebases without manual explanation.
SquarePact Word Add-In Connects to OpenClaw Gateway for Agentic Document Editing
SquarePact's Word add-in connects to your own OpenClaw gateway over websocket. The agent proposes edits shown as approval cards, with stale edit detection and retry logic for malformed JSON from smaller models.
SkillOpt: Optimizing Markdown Skill Files as Trainable Parameters for AI Agents
SkillOpt formalizes the ad-hoc process of editing markdown skill files for AI coding agents, using frontier models to propose bounded edits gated against validation sets. Best skills converge with 1-4 accepted edits out of many proposals, and transfer across models like Codex to Claude Code.
Eden AI: European API Hub for AI Models – Pivots as OpenRouter Alternative
Eden AI offers a single unified API to access 500+ AI models (LLMs, vision, OCR, speech) with smart routing, fallback mechanisms, and region control. Positioned as a European alternative to OpenRouter.