KADATH
Evolutionary multi-agent runtime that breeds, evaluates, and improves autonomous agents across reproducible epochs to converge on optimization of a goal.
What is KADATH?
Evolutionary multi-agent runtime that breeds, evaluates, and improves autonomous agents across reproducible epochs to converge on optimization of a goal.
Overview of KADATH
What it is
KADATH (Kernel for Agentic Darwinian Adaptation, Tooling, and Heredity) is an evolutionary multi-agent runtime. It breeds, evaluates, and improves autonomous agents across reproducible epochs to converge on optimization of a goal. The system uses repeated competition and selection to produce the best agent for a user's intended outcome.
Key features
- Evolves complete agent frameworks including system prompt, Python implementation, tools, dependency declarations, and supporting files
- Two distinct layers: a non-evolving kernel that controls the run, benchmark, deadlines, containers, grading, and lineage; and evolvable organisms that perform the goal
- Organism genomes are mounted read-only during an epoch; changes only occur during the post-grade mutation phase
- Supports a machine-readable benchmark contract with finite score ranges, weighted rubrics, automatic failure rules, and anti-fraud checks
- Provides a live dashboard showing epoch progress, leaderboard, model-call count, worker activity, crash count, and recent high-level activity
- Retains operational memory, detailed evidence traces, and model request/response logs with secret redaction
Use cases
KADATH is designed for scenarios where a user wants to optimize an autonomous agent's performance on a specific goal through evolutionary competition. It can be used to iteratively improve agent behavior across multiple generations, with each generation evaluated against a locked benchmark.
Getting started
Launching ./kadath.sh opens the interactive terminal interface. On first launch, KADATH asks for an OpenAI API key and the OpenAI model ID. It generates PostgreSQL, MinIO, LiteLLM, and SearXNG secrets locally. The configuration is stored in .kadath/config.env. For a new run, the interface collects the goal, epoch duration, population size, and number of epochs.
License and pricing
License: Apache-2.0. Pricing: open source.
這是什麼
KADATH(Kernel for Agentic Darwinian Adaptation, Tooling, and Heredity)是一個演化式多代理執行環境。它在可重現的世代中培育、評估並改進自主代理,以收斂於目標的最佳化。系統透過重複的競爭與選擇,為使用者預期的結果產生最佳代理。
主要功能
- 演化完整的代理框架,包括系統提示詞、Python 實作、工具、依賴宣告與支援檔案
- 兩個不同的層級:不演化的核心控制執行、基準測試、截止時間、容器、評分與譜系;可演化的有機體執行目標
- 有機體基因組在世代期間以唯讀方式掛載;僅在評分後的突變階段才能變更
- 支援機器可讀的基準測試合約,包含有限分數範圍、加權評分標準、自動失敗規則與反詐欺檢查
- 提供即時儀表板,顯示世代進度、排行榜、模型呼叫次數、工作者活動、當機次數與近期高層級活動
- 保留操作記憶、詳細證據追蹤記錄與模型請求/回應日誌,並對機密資訊進行遮罩處理
使用情境
KADATH 適用於使用者希望透過演化競爭來最佳化自主代理在特定目標上表現的情境。它可用於在多個世代中反覆改進代理行為,每個世代都根據鎖定的基準測試進行評估。
快速開始
執行 ./kadath.sh 會開啟互動式終端介面。首次啟動時,KADATH 會要求提供 OpenAI API 金鑰與 OpenAI 模型 ID。它會在本機產生 PostgreSQL、MinIO、LiteLLM 與 SearXNG 的機密資訊。設定檔儲存在 .kadath/config.env。若要開始新的執行,介面會收集目標、世代持續時間、群體大小與世代數量。
授權與價格
授權:Apache-2.0。價格:開放原始碼。
Based on the project's public README and repository metadata.