# Adala

> Adala is an open-source Autonomous Data (Labeling) Agent framework that enables LLM-powered agents to independently acquire and apply data processing skills through iterative learning.

Adala is an open-source Python framework built by HumanSignal for creating autonomous agents specialized in data labeling and processing tasks. Released under the Apache License 2.0, it is freely available on GitHub and installable via pip. The framework connects LLMs (such as OpenAI's GPT models or any OpenAI-compatible API) to structured datasets, enabling agents to iteratively learn and refine skills without manual intervention.

## What It Is

Adala stands for **A**utonomous **DA**ta (**L**abeling) **A**gent framework. At its core, it provides a structured way to define agents that can independently develop data-processing skills by learning from a ground truth dataset. The agent's learning process is shaped by its environment (the dataset), its observations (model outputs), and its reflections (feedback loops). The framework treats LLMs as "runtimes" — the execution engine where skills are applied — making it straightforward to swap between different model providers.

## Core Architecture

Adala is organized around five primary abstractions:

- **Agents** — the main interface that orchestrates learning and inference
- **Environments** — where ground truth data is provided (e.g., `StaticEnvironment` wrapping a pandas DataFrame)
- **Skills** — discrete, configurable capabilities such as classification, summarization, translation, question answering, and text generation
- **Runtimes** — the LLM backends that execute skills (supports OpenAI, OpenRouter, and any OpenAI-compatible endpoint including Claude and Gemini via OpenRouter)
- **Memories** — storage for agent state and learned knowledge across iterations

## Skill Types and Examples

The repository ships with a range of ready-to-use skill implementations and Colab-runnable notebooks:

- **ClassificationSkill** — label text into predefined categories, with optional Chain-of-Thought reasoning
- **SummarizationSkill** — condense text into shorter form
- **QuestionAnsweringSkill** — answer questions from context
- **TranslationSkill** — translate between languages
- **TextGenerationSkill** — generate text from prompts
- **OntologyCreator** — infer ontologies from text examples
- **Skill sets** — chain multiple skills in sequence for complex pipelines

## Autonomous Learning Model

What distinguishes Adala from simple LLM wrappers is its iterative self-improvement loop. When `agent.learn()` is called, the agent runs multiple learning iterations against a labeled training dataset, comparing its outputs to ground truth and refining its skill instructions until a configurable accuracy threshold is met. This student/teacher architecture allows a lower-cost model to handle inference while a more capable model guides instruction refinement — a pattern the framework explicitly supports through separate `default_runtime` and `teacher_runtimes` configurations.

## Audience and Use Cases

The README identifies four primary audiences: AI engineers building production agent pipelines, machine learning researchers experimenting with problem decomposition, data scientists preprocessing and postprocessing large DataFrames in Python notebooks, and educators or students exploring agent-based AI. The framework is designed to abstract low-level ML complexity while remaining extensible for advanced customization.

## Current Status

The latest tagged release is **v0.0.4** ("Introducing a New Code Environment and Enhanced Automated Prompt Engineering"), published in November 2023. The GitHub repository shows ongoing activity with the last push recorded in July 2026 and 169 open issues, indicating active development beyond the versioned release. The project roadmap lists planned features including multi-task learning, Named Entity Recognition skills, a REST API, command-line utilities, and vision/multi-modal agent skills. The repository has accumulated over 1,600 stars and 156 forks on GitHub.

## Features
- Autonomous agent learning from ground truth datasets
- Iterative skill refinement with configurable accuracy thresholds
- ClassificationSkill with optional Chain-of-Thought reasoning
- SummarizationSkill for text condensation
- TranslationSkill for multilingual processing
- QuestionAnsweringSkill from context
- TextGenerationSkill from prompts
- OntologyCreator for inferring ontologies from text
- Skill sets for chaining multiple skills in sequence
- Student/teacher runtime architecture for cost-efficient learning
- Support for OpenAI, OpenRouter, Claude, Gemini, and OpenAI-compatible LLMs
- StaticEnvironment for pandas DataFrame integration
- Modular Agents, Environments, Memories, Runtimes, and Skills abstractions
- Python notebook-native usage
- Apache 2.0 open-source license

## Integrations
OpenAI GPT-4o, OpenAI GPT-3.5-turbo, OpenRouter, Anthropic Claude 3.5 Haiku, Google Gemini, pandas, Label Studio

## Platforms
CLI, API, DEVELOPER_SDK

## Pricing
Open Source

## Version
0.0.4

## Links
- Website: https://humansignal.github.io/Adala/
- Documentation: https://humansignal.github.io/Adala/
- Repository: https://github.com/HumanSignal/Adala
- EveryDev.ai: https://www.everydev.ai/tools/adala