Self-Evolving Agent Harness (Harness-RSI): Formalizing Fixed-Weight Agent Scaffolding Evolution with Continual PAC-Learning Bounds

Background and the Problem

In production agent deployments, foundation model weights are typically frozen—either behind proprietary APIs or due to the prohibitive cost of continuous fine-tuning. Consequently, practical capability gains rely heavily on evolving the agent's software scaffolding (Agent Harness), including system prompts, memory schemas, workflow graphs, executable tools, and plugin compositions. However, ad-hoc automated harness self-modification frequently suffers from silent regressions on historical tasks, compounding errors, and runtime privilege risks due to the lack of formal verification guarantees.

Architecture and How It Works

To establish a rigorous foundation for weight-frozen self-improvement, researchers introduced the Self-Evolving Agent Harness (Harness-RSI) framework:

1. Decoupling Update Targets from Proposal & Admission: Separates modifications to interpreted state (prompts, episodic memory) from structural revisions to executable tools, repository-level control logic, and plugin orchestration;
2. Continual PAC-Learning Generalization & Retention Bounds: Formulates harness search via evolutionary optimization analyzed through continual PAC-learning, deriving expected generalization bounds, retention bounds conditioned on historical regression budgets, and a high-probability admission rule evaluated on fresh independent validation data;
3. Tiered Safeguards for Persistent Revisions: Establishes independent isolation layers governing autonomous learning, runtime execution privileges, and consequential external actions.

Benchmarks and Measured Results

Evaluations and cross-system case studies across model-harness co-design and plugin reuse demonstrate:

• Bounded Historical Regression: Enforcing high-probability admission rules with explicit regression budgets prevents iterative harness drift while steadily accumulating verified skills;
• Scalable Plugin & Workflow Reuse: Persistent validated revisions accelerate multi-agent organizational workflows without requiring gradient updates to the underlying LLM.

Getting Started for Developers

The 66-page paper, algorithms, and reference implementations are available at https://github.com/zgcagi/Harness-RSI:

• Implement PAC-Gated Harness Commits: Before persisting agent-proposed edits to system prompts, skills, or tool scripts, evaluate the candidate harness against a held-out validation suite and enforce a strict regression budget;
• Separate Interpreted State from Control Logic: Allow frequent autonomous updates to episodic memory while gating tool-code mutations behind deterministic sandbox verification.