PrimeIntellect-ai/prime-agent

A self-improving RLM agent for coding workflows and long-running autonomous tasks.

What it solves

Prime Agent is designed for complex, long-running coding and research tasks that typically exceed the limits of a single chat session. It solves the problem of losing context, state, and progress when a terminal disconnects or when a task requires parallel execution and self-correction over time.

How it works

The project uses two primary abstractions: the Recursive Language Model (RLM), which treats prompts as variables and tools as function calls within a persistent Python REPL, and a Continual Harness that stores durable state like memories, skill descriptions, and subagent specifications. This allows the agent to refine its own operating patterns through evidence-backed updates via the /refine command.

Who it’s for

It is primarily aimed at developers and AI researchers who need an autonomous agent capable of handling extended workflows, managing subagents, and executing code in a persistent environment.

Highlights

  • Programmatic Control: Uses a persistent Python REPL for file operations, shell commands, and context management.
  • Subagent Orchestration: Built-in rlm() function to spawn child agents for parallel or background work.
  • Self-Improving Harness: The /refine command allows the agent to update its supplemental state based on lessons learned during a session.
  • Daemon-backed Continuity: Sessions run as background services, allowing users to detach and reattach to active agents.
  • Autonomous Mode: Supports bounded autonomy with configured budgets for turns, tokens, and time, including user-defined quality gates.
  • Executable Skills: Recurring workflows can be turned into importable Python packages.

Related

  • Dispatch
    Prime Agent self-improving RLM harness release

    Prime Agent is an open‑source, self‑improving coding harness built on Recursive Language Models and a continual‑state harness, achieving state‑of‑the‑art scores on ARC‑AGI‑3 and competitive performance on long‑context benchmarks.

  • Project
    greyhaven-ai/autocontext

    A recursive self-improving harness that helps AI agents iteratively solve tasks by running evaluations, retaining successful lessons, and distilling stable behavior.

  • Project
    avbiswas/fast-rlm

    A minimal implementation of Recursive Language Models (RLMs) that allows LLMs to handle arbitrarily long prompts by programmatically exploring data via a Python REPL and recursive sub-agents.

  • Project
    letta-ai/letta-code

    A stateful agent harness that enables AI agents to maintain long-term memory, identity, and the ability to self-evolve through programmatic context rewriting.