File: docs_site/docs/.vuepress/dist/assets/basic-concepts-Bz68xxmp.js

import{A as e,d as t,m as n,t as r}from"./plugin-vue_export-helper-DazhzKqf.js";var i=JSON.parse(`{"path":"/guide/basic-concepts.html","title":"Basic Concepts","lang":"en-US","frontmatter":{},"git":{},"filePathRelative":"guide/basic-concepts.md"}`),a={name:`basic-concepts.md`};function o(r,i,a,o,s,c){return e(),t(`div`,null,[...i[0]||=[n(`<h1 id="basic-concepts" tabindex="-1"><a class="header-anchor" href="#basic-concepts"><span>Basic Concepts</span></a></h1><p>HyperFlow is a self-improving agent framework. Instead of manually tuning an AI agent, you let another AI agent do it automatically.</p><p>The core idea comes from evolutionary computation and Quality-Diversity (QD) algorithms: maintain an <strong>archive</strong> of agent versions, evaluate each one, and use the best-performing versions as starting points for further improvement.</p><h2 id="the-two-agents" tabindex="-1"><a class="header-anchor" href="#the-two-agents"><span>The Two Agents</span></a></h2><h3 id="taskagent-?-the-worker" tabindex="-1"><a class="header-anchor" href="#taskagent-?-the-worker"><span>TaskAgent ? The Worker</span></a></h3><p>The TaskAgent serves as the problem solver for domain-specific tasks. It receives a formatted prompt, optionally uses tools, and returns a prediction.</p><ul><li><strong>Input</strong>: A task description.</li><li><strong>Output</strong>: A prediction.</li><li><strong>Tools</strong>: Domain-specific, optional.</li></ul><h3 id="metaagent-?-the-improver" tabindex="-1"><a class="header-anchor" href="#metaagent-?-the-improver"><span>MetaAgent ? The Improver</span></a></h3><p>The MetaAgent&#39;s job is to make the TaskAgent better. It acts as the &quot;mutation operator&quot;. It reads past evaluation results, identifies failures, and edits source code (prompts, logic, tools) to fix them.</p><ul><li><strong>Input</strong>: Repo path + eval results path + parent score</li><li><strong>Output</strong>: Modified source code on disk (patches/diffs)</li><li><strong>Tools</strong>: <code>bash</code> (run shell commands) + <code>editor</code> (view/edit files) ? built-in</li></ul><h2 id="the-evolutionary-loop" tabindex="-1"><a class="header-anchor" href="#the-evolutionary-loop"><span>The Evolutionary Loop</span></a></h2><p>The evolutionary loop (<code>run_generate_loop</code>) is the heart of the system. It runs multiple generations, each improving on a previous one.</p><div class="mermaid"> graph TD A[Archive] --&gt;|1. Select Parent| B(Setup Executor) B --&gt;|2. Apply Lineage Patches| C{MetaAgent Runs} C --&gt;|3. Reads failures &amp; modifies code| D[TaskAgent Runs] D --&gt;|4. Evaluated against harness| E{Harness Evaluation} E --&gt;|5. Store score &amp; diff| A style A fill:#2a9d8f,stroke:#fff,stroke-width:2px,color:#fff style C fill:#e76f51,stroke:#fff,stroke-width:2px,color:#fff style D fill:#e9c46a,stroke:#fff,stroke-width:2px,color:#333 </div><h2 id="the-archive" tabindex="-1"><a class="header-anchor" href="#the-archive"><span>The Archive</span></a></h2><p>The archive is the central data structure stored as a JSON Lines (JSONL) file that records the history of all generations. It serves as a <strong>versioned record of evolutionary improvement</strong>, acting as a family tree where different generations can branch from different original parent versions.</p>`,15)]])}var s=r(a,[[`render`,o]]);export{i as _pageData,s as default};