Adala (Autonomous DAta Labeling Agent) is an open-source Python framework developed by HumanSignal, the company behind Label Studio. It enables developers to build autonomous agents that independently learn data processing skills—such as classification, summarization, and generation—through iterative feedback loops powered by large language models. The framework is grounded in a student/teacher architecture where agents operate within a configurable 'runtime' environment (an LLM provider such as OpenAI or VertexAI). Agents start from a set of ground truth examples, apply skills to new data, receive feedback on errors, and iteratively refine their prompts until accuracy meets a defined threshold. This design reduces the need for continuous human annotation while maintaining quality control. Adala exposes three core skill primitives: TransformSkill (modify or label records), AnalysisSkill (derive insights), and SynthesisSkill (generate new data). Skills can be composed into sequential or parallel SkillSets, and a CodeEnvironment feature allows agents to generate, execute, and validate Python code as part of their workflow. The framework is aimed at AI engineers, ML researchers, and data scientists who need to automate data labeling pipelines at scale. Because it is LLM-agnostic, users can swap runtime providers, and the student/teacher pattern can reduce per-prediction inference costs significantly compared to always using a large frontier model. As of June 2026 the project remains in early/experimental release (latest tagged version 0.0.4) and is not yet recommended for production use, though active development and community contributions continue on GitHub. Key features: - Autonomous skill acquisition: agents iteratively learn labeling skills from ground truth examples via LLM feedback - Student/teacher architecture: a smaller student model learns from a larger teacher model, reducing inference costs - Composable skill primitives: TransformSkill, AnalysisSkill, SynthesisSkill, with sequential and parallel SkillSets - CodeEnvironment: agents can generate, execute, and validate Python code as part of task execution - LLM-agnostic runtime: supports OpenAI, VertexAI, and other providers via a swappable runtime interface - Ground-truth-anchored training: skill development is anchored to verified labeled examples for reliability - Open-source and free: MIT-licensed, hosted on GitHub under HumanSignal organization - Integration with Label Studio: designed to complement HumanSignal's human-in-the-loop labeling platform
Free and open-source (MIT license). Users pay only for LLM API usage (e.g., OpenAI tokens) during training and inference.