Course contentsShow
← Machine Learning and Deep Learning
Lesson 2062 of 3,53844. AI Agents: FoundationsFree lesson

Action Space and Tool Registry

Define the set of available actions or tools an agent can invoke and how they're registered and described.

Action Space and Tool Registry

What you'll learn: How AI agents define, organize, and access the set of tools they can use to accomplish tasks.


The Core Idea

An action space is the complete set of operations an agent can perform—its "toolbox." A tool registry is the system that catalogs these tools with structured descriptions so the agent knows what's available and how to use each one.

Think of it like a restaurant kitchen: the action space is all the cooking techniques a chef can use (sauté, grill, bake), while the tool registry is the recipe book that describes each technique, what ingredients it needs, and what it produces.

Why This Matters

When implementing the ReAct pattern (from our previous lesson), the agent must choose which action to take. But how does it know what actions exist? The tool registry provides:

  1. Tool names – identifiable labels like search_web or calculate
  2. Descriptions – what each tool does in natural language
  3. Input schemas – what parameters each tool requires
  4. Return types – what kind of output to expect

The LLM reads these descriptions to decide which tool fits its reasoning step, then the agent runtime invokes that tool with the appropriate parameters.

Structured Registration

Tools are typically registered with metadata:

  • Name: get_weather
  • Description: "Fetches current weather for a given city"
  • Parameters: {"city": "string", "units": "string (optional)"}
  • Returns: Weather data object

This structured format lets the LLM generate valid tool calls, and the agent framework routes those calls to the correct implementation.

Dynamic vs. Static Action Spaces

Some agents have a fixed set of tools (defined at initialization), while agentic RAG systems may dynamically add or remove tools based on the task context—like loading domain-specific calculators only when needed.

Key Takeaway: The action space defines what an agent can do, and the tool registry provides the structured metadata that enables the agent to discover and invoke those capabilities intelligently.