zgcagi/Harness-RSI

What it solves

Large language models are usually treated as fixed: you don't retrain them for each new task. But the "harness" around a model — the tools it can call, the prompts it reads, its memory, and how its execution is controlled — can be updated. This project surveys and organizes the growing body of research on harness self-evolution: systems where the agent's own execution experience is used to revise that harness state, so later runs perform better under the same base model. It addresses the lack of a unified picture of this fast-moving area by providing a definition, a taxonomy, and a categorized reading list.

How it works

A harness H is defined as the software and configuration governing how a model M operates in an environment E — including tools, context construction, memory, and execution control. An evaluator J scores each trajectory, and harness self-evolution uses a batch of trajectories to compute an updated harness for later runs, via an update mechanism that is itself available for further revisions. Accepted states are saved and loaded later; rejected candidates leave the harness unchanged. The survey keeps the base model fixed throughout.

The resource organizes the field into two main update-target branches plus separate branches for evaluation and theory:

• Interpreted updates revise content read by a fixed host: prompts and instructions (scalar-score search, textual-gradient critique, self-referential mutation), memory and reusable procedures (experience mining, procedural skillbooks, gated consolidation), and workflows and topologies (fixed-pipeline compilation, graph/topology search, feedback-driven workflow revision).
• Executable updates revise the implementations the host actually invokes: callable tool synthesis and evidence-grounded component patching.

It also lists precursors (e.g., EURISKO, Gödel Machines), a set of related surveys, and companion materials including the full 66-page report, a short 18-page paper, taxonomy data, and a project page.

Who it's for

Researchers, AI engineers, and graduate students working on agentic systems, self-improving agents, prompt and skill optimization, and LLM application development. The resource is structured as a reading guide with topic-indexed entries (each with paper links), so it suits anyone wanting to map the landscape of self-evolving harnesses, find relevant prior work, or locate the authors' own survey and taxonomy.

Highlights

• Defines harness self-evolution and distinguishes it from mere self-improvement (self-improvement requires a beneficial change established through evaluation).
• Provides a taxonomy splitting update targets into interpreted (prompts, memory, workflows) and executable (tools, code patches) categories, with evaluation and theory as separate branches.
• Curates a large, categorized literature list — from APE and Reflexion to DSPy, Voyager, and beyond — each with links to papers and, where available, code and projects.
• Includes a 66-page full survey report, an 18-page short paper, taxonomy data (JSON), and a set of related surveys for orientation.
• Covers historical precursors such as EURISKO and Gödel Machines to ground the field's lineage.