Basic Concepts

This page briefly introduces the core concepts used across our community docs: Agent, Self-Evolution (RSI), Harness, Harness-Level vs. Model-Level Self-Evolution, and Benchmark.

What is an Agent?

An Agent is a system powered by a foundation model that can perceive its environment, plan autonomously, invoke tools (editing files, running commands, browsing the web, etc.), and iteratively complete complex tasks based on environment feedback.

Unlike a single-turn chat model ("text in → text out"), an Agent operates in a multi-step closed loop (Observe → Think → Act → Feedback) to deliver verifiable artifacts.

What is Self-Evolution (RSI)?

Self-Evolution (Recursive Self-Improvement, RSI) is the process where an agent uses its own execution trajectories (Traces) and evaluation feedback to automatically diagnose weaknesses and update its configuration or model weights, continuously improving on future tasks.

A standard loop has four steps: Forward Rollout → Critique → Update → Verify & Select.

1. Rollout: Generate execution traces.
2. Critique: Score artifacts and diagnose weaknesses.
3. Update: Evolve Harness components or model weights.
4. Select: Retain improved candidates for the next generation.

What is a Harness?

If the foundation model is the "brain", the Harness is the runtime architecture and control layer wrapped around the model that turns it into a reliable engineering system (such as PenguinHarness). It includes:

• System Prompts: e.g., AGENTS.md, defining operating rules and workflows.
• Skills: e.g., SKILL.md and scripts, packaging reusable domain knowledge and procedures loaded on demand.
• Tools & Sandbox: Execution interfaces for files, shell commands, browsers, and MCP servers.
• Memory & Orchestration: Cross-session memory, context compaction, and multi-agent (subagent) coordination.

1. Outer Harness Layer (Stage 1 Evolution Target): Prompts (AGENTS.md), Skill Library (Skills), Tools & Sandbox (Tools & MCP), and Memory.
2. Inner Model Layer (Stage 2 Evolution Target): Foundation Model weights evolved via Trajectory Distillation, SFT, and Agentic RL.

Harness-Level vs. Model-Level Self-Evolution

Depending on what gets updated, self-evolution falls into Harness-Level Self-Evolution (evolving outer components) and Model-Level Self-Evolution (evolving model weights):

Dimension	Harness-Level Evolution (Stage 1)	Model-Level Evolution (Stage 2)
Target	Prompts, Skills, tool code, memory, workflows	Neural network weights (full weights or LoRA)
Methods	Prompt optimization, skill synthesis/pruning, reflection	Trajectory distillation, SFT, Agentic RL
Compute Barrier	Low: Inference only (API calls); no training cluster needed	High: Multi-GPU clusters for rollouts and gradient updates
Iteration Speed	Low cost, minutes to hours	High cost, days to weeks
Interpretability	High: Human-readable Markdown/code, easy to audit and roll back	Black-box: Encoded in dense floating-point weights
Transferability	Strong: Evolved skills and workflows transfer across models	Model-specific: Tied to a single base model

Our Roadmap progresses from Harness-Level Evolution to Model-Level Evolution: first accumulating benchmarks, algorithms, and high-quality traces at low cost, then distilling and reinforcing that experience into model weights.

What is a Benchmark?

In self-evolution research, a Benchmark is a standardized task suite and automated grading environment that drives evolution and measures generalization (see RFC-0001):

• Cases: Each case provides an initial workspace (input_dir) and a task prompt (Prompt).
• Evaluator: Automated scripts, unit tests, or LLM judges that score final artifacts in [0, 1].
• Splits: Partitioned into Training (train, for evolving the agent), Validation (val, for selection and preventing overfitting), and Test (test, strictly frozen for final evaluation).

Together, Benchmarks provide the arena, Algorithm Plugins provide the evolution algorithms, and Evaluations record pre- and post-evolution scores and traces:

1. Benchmark: Provides train / val / test splits and automated grading.
2. Algorithm Plugin: Drives agent self-evolution on training splits.
3. Evaluation: Records scores, cost, and traces for the leaderboard.