## Agent Architecture
Agents are built on several key components that work together to create specialized AI assistants:
A clear, descriptive identifier that indicates the agent's purpose and makes it easy to find and select when starting conversations.
The **system prompt** is the foundational set of instructions that defines your agent's identity and behavior. Unlike regular conversation messages that can be modified or ignored, the system prompt is persistent and authoritative. This prompt shapes every response the agent generates.
**System prompts include:**
* Role definition and expertise areas
* Communication tone and style
* Operational rules and constraints
* Business context and policies
* Decision-making frameworks
* Error handling procedures
The quality of your system prompt directly impacts agent effectiveness. Well-crafted instructions produce consistent, reliable results; vague instructions lead to unpredictable behavior.
The AI model is an important choice that significantly affects both performance and cost. Different models have different capabilities, speed, and pricing.
**Key considerations:**
* **Performance**: More capable models (thinking models) excel at complex reasoning but are more expensive
* **Cost**: Simpler models are cost-effective for straightforward tasks
* **Speed**: Faster models reduce latency for high-frequency interactions
* **Context window**: Determines how much information the agent can process at once
Model selection impacts your agent's quality and operational costs. Match model capability to task complexity: don't overspend on simple tasks or underinvest in complex workflows.
In addition to the model choice, you can adjust **model parameters** that control response randomness, creativity, and repetition. See [Creating Your First Agent](/ai-agents/creating-your-first-agent) for detailed parameter configuration.
Agents can be equipped with various capabilities beyond text generation:
**Code Execution**: Run code snippets to perform calculations, data transformations, or complex logic operations.
**Web Search**: Access current information from the internet when needed for real-time data or recent events.
**External Connections (MCP)**: The most important capability, direct integration with your business systems through the Model Context Protocol.
Agents dynamically select which tools to use based on the task and their system prompt guidance.
**MCP** connects agents to your business systems for real actions.\
See [Tools Connection](/tools-connection/introduction) for setup details.
## Execution Loop
Agents operate autonomously through a continuous loop, iterating until they achieve the desired result:
```mermaid theme={null}
graph TD
A[User Input] --> B[Analyze & Plan]
B --> C[Select Approach]
C --> D[Execute Actions]
D --> E{Evaluate Results}
E -->|Success| F[Return Final Result]
E -->|Needs Refinement| G[Try Different Tools/Approach]
G --> C
style A fill:#4A90E2,stroke:#2E5C8A,stroke-width:2px,color:#fff
style B fill:#7B68EE,stroke:#5A4DB5,stroke-width:2px,color:#fff
style C fill:#7B68EE,stroke:#5A4DB5,stroke-width:2px,color:#fff
style D fill:#7B68EE,stroke:#5A4DB5,stroke-width:2px,color:#fff
style E fill:#FF6B6B,stroke:#CC5555,stroke-width:3px,color:#fff
style F fill:#51CF66,stroke:#40A852,stroke-width:2px,color:#fff
style G fill:#FFA94D,stroke:#CC8539,stroke-width:2px,color:#fff
```
Agents iterate through multiple approaches and tool combinations until they successfully complete the task or reach defined limits.
## Model Selection
WonkaChat supports multiple AI providers, allowing you to choose any model that best fits your needs. Model selection depends on several factors:
* **Task complexity**: More complex reasoning requires more capable models
* **Response speed**: Some models prioritize faster responses
* **Cost considerations**: Model pricing varies significantly across providers
For detailed guidance on selecting the right model for your agent, see [Creating Your First Agent](/ai-agents/creating-your-first-agent).
## Security and Access Control
### Permission Inheritance
Agents operate within your existing security boundaries. They cannot perform any action you couldn't perform yourself without WonkaChat. If you lack permission to access a system, read a file, or execute an operation manually, the agent cannot bypass those restrictions for you.
See [Security & Governance](/security-governance/access-control) for comprehensive security details.