Skip to main content

Agentic Framework

The aiXplain agentic framework enables you to build trusted, modular, and adaptable agentic AI solutions. It's designed for production use while remaining accessible and flexible enough to support rapid prototyping.

Each agent is composed of well-defined modular components that communicate through structured data interfaces:

  • Models: Large language models (LLMs) that interpret instructions, select tools, review outcomes, and adapt execution plans.
  • Instructions: Natural language prompts that guide agent behavior and response style.
  • Tools: A collection of capabilities—including AI assets, custom Python functions, pipelines, data sources, and other agents—accessed through a unified interface to perform specific tasks.
  • Memory: A shared store used to track instructions, intermediate outputs, and task state, while also providing persistent storage for user context and multi-turn interactions across a session.
  • Guardrails (coming soon): Policies and filters that enforce content safety, access control, compliance, and input/output validation.

These components allow you to create agentic solutions that solve complex problems and automate execution through AI reasoning. Tracing and logging allow you to see under the hood—every thought and step the agent takes—for debugging and iteration, as well as for auditing and tracking guardrail violations.

Docusaurus themed imageDocusaurus themed image

Agent design patterns​

aiXplain supports several reusable agent design patterns, optimized for different classes of problems:

Autonomous agent​

A general-purpose agent that interprets instructions and adapts behavior dynamically using tools and memory. Ideal for flexible decision-making and natural language interfaces.

RAG agent (retrieval-augmented generation)​

Retrieves relevant context from external data sources—such as online search, structured datasets, or vector databases—before generating a response.

Workflow agent​

Incorporates pipelines as agent tools to ensure deterministic execution of structured tasks. Suitable for repeatable, multi-step processes that follow a defined path.

Team agent​

A role-based multi-agent orchestrator that distributes responsibilities across built-in specialized agents and treats agents as tools to handle multi-step execution:

  • Mentalist agent: Plans and decomposes the task
  • Inspector agent: Applies guardrails on prompts and responses to validate safety, quality, cost, and performance
  • Bodyguard agent: Applies role-based access control for data access and privacy
  • Orchestrator: Oversees function calling and coordination
  • Response agent: Generates task-specific outputs or insights

Evolver agent (coming soon): Attaches to any agent to observe its behavior, iterate on its instructions and tools, and perform continuous benchmarking to suggest improvements.

This ensemble is designed for specialization and cross-validation, enabling collaborative problem solving, policy enforcement, and robust task execution.

Get started​

Ready to build your first agent? Check out the Quickstart guide to begin creating and deploying agents using the aiXplain framework.