AetherLabsAI/RSIAgent

What it solves

RSIAgent addresses the challenge of enabling AI agents to adapt to new digital environments (like complex software) without needing to update the underlying model's weights. It provides a way for agents to autonomously explore a system, learn how it works, and store that knowledge for later use in specific tasks.

How it works

The framework uses a "broad-then-deep" exploration strategy coordinated by three specialized agents:

• Curriculum Agent: Selects informative tasks to explore and decides when further practice is necessary.
• Actor Agent: Interacts with the environment using Python or Bash scripts and visual observations. It distills its experiences into a persistent memory.
• Verifier Agent: Independently checks the results of the Actor's actions against the task requirements to provide grounded feedback.

Learning happens in three phases:

1. Broad Recursive Self-exploration (BRS): Parallel exploration of diverse projects to build a general knowledge base.
2. Deep Recursive Self-exploration (DRS): Focused practice on hard cases and boundary conditions to refine memory.
3. Test-time memory reuse: The agent uses the frozen memory to execute downstream tasks without further learning.

Who it’s for

It is designed for developers and researchers working on autonomous agents, specifically those focusing on computer-use benchmarks and recursive self-improvement in software environments.

Highlights

• Training-free: Improves performance without updating model parameters.
• Recursive Learning: Uses a feedback loop of exploration, verification, and memory consolidation.
• Persistent Memory: Stores procedures, scripts, and lessons from failures for reuse.
• Broad-then-Deep Strategy: Moves from diverse experience to targeted investigation of fragile successes.