Skip to content

Core Concepts

Understanding the foundational elements of Minions is essential before diving into its deeper architecture. This section introduces the two most important constructs: Agents and Recipes.

What is an Agent?

In Minions, an Agent is an autonomous entity capable of executing a goal-driven process. It engages in multi-step reasoning, decision-making, and tool interaction. Each agent operates within a defined context, maintaining state across steps and coordinating its behavior using recipes, memory, and step graphs.

Agent Lifecycle

  1. Initialization: The agent is instantiated from a recipe. This includes loading its system prompt, goal, memory configuration, and step graph.

  2. Goal Assignment: A goal is defined either at instantiation or injected dynamically (e.g., from user input or a higher-level planner).

  3. Step Execution: The agent traverses its step graph, making decisions, calling models, and interacting with tools as needed.

  4. Memory Management: The agent maintains multiple memory subsystems:

    • Episodic memory for chronological event storage
    • Vector memory for semantic similarity search
    • Short-term memory for active context
    • Long-term memory for persistent knowledge

  5. Memory Summarization: After each step, the agent:

    • Queries relevant memory based on context
    • Generates summaries using model calls
    • Updates appropriate memory subsystems
    • Maintains memory efficiency through summarization

  6. Completion: The agent completes once the graph reaches a terminal node or a completion condition is met.

Agents are MCP-compliant and traceable, allowing developers to monitor execution, review decision points, and integrate with observability tooling.

What is a Recipe?

A Recipe is a declarative blueprint for constructing an agent. Rather than configuring agents imperatively through code, developers define recipes that describe the agent's identity, structure, and behavior.

A recipe includes:

  • System Prompt – The initial set of instructions or framing context for the agent's behavior.

  • Goal – The task or objective the agent is trying to achieve.

  • Step Graph – A flow structure of logical steps the agent executes in order, conditionally, or iteratively.

  • Memory Configuration – Defines how the agent manages memory:

    • Memory subsystem selection and configuration
    • Query strategy configuration
    • Summarization settings
    • Persistence strategy selection

  • Toolchain and Hooks (optional) – Tools the agent can call and hooks for evaluation, logging, or transformation.

By encapsulating these into a recipe, Minions makes it easy to spin up different kinds of agents, each purpose-built for a use case, without rewriting orchestration logic.