Autonomous Agents & Dynamic Tool Registry
This feature details agent personality configuration, prompt templating, and dynamic tool execution.
1. Agent Personality & Specification (AiAgentSpec)
Each AiAgent CR defines a declarative agent configuration:
- System Prompt: Core persona definition.
- User Prompt Template: Default query template supporting {{input}} parameter substitution.
- Provider Reference: Links to LlmProvider CR (providerRef.name).
- Model Override: Optional agent-specific model override (e.g. gpt-4o-mini).
- Tools: List of active capabilities enabled for this agent.
- Guardrails: Safety policy definitions for prompt sanitization and output validation.
2. Built-in Tools & Dynamic Tool Registry (ToolRegistry)
Tools provide capabilities that agents can execute during execution pipelines:
| Tool | Description | Usage Example |
|---|---|---|
calculator |
Evaluates basic mathematical and financial calculations (+, -, *, /). |
"Calculate total order cost with 18% VAT" |
web-search |
Performs real-time web search for technical documentation. | "Search latest Spring Boot 4.1 release notes" |
weather |
Retrieves real-time simulated weather data for specified locations. | "Check current weather in Istanbul" |
2.1 Virtual Thread Tool Execution
Active tools execute concurrently on Java 25 Virtual Threads via ToolRegistry.executeActiveTools(), collecting outputs and injecting them into the LLM system prompt context before inference.
3. Dynamic Tool Loading (toolSources)
Agents can load custom tools at runtime without redeploying operator code:
- CONFIGMAP: Load JavaScript/Python tool scripts from Kubernetes ConfigMaps.
- JAR: Load compiled Java tool plugins dynamically from a shared volume or object storage.
- FOLDER: Watch local filesystem directories for hot-reloadable tool definitions.