Agent Architect

Design your own AI agents — choose LLM, define prompt chains, and wire them into any workflow node.

Overview

Agent Architect is a visual designer for creating custom AI agents from scratch or from a pre-built template. You define every aspect of the agent: which language model it uses, what its system prompt says, which tools from the Tool Registry it can call, and which knowledge bases it can query.

Once saved, a custom agent appears as a draggable node type in the Workflow Editor — identical in behaviour to the pre-built agent types but with your configuration baked in.

Agent Configuration

Model

Select the language model to back this agent. Models available in the dropdown are those currently active in the Model Center. You can choose cloud models (Claude, GPT-4o) or local models (TinyLlama, Mistral) depending on your setup.

System Prompt

The system prompt defines the agent's identity, task scope, and output format. Write a clear, specific system prompt to get consistent behaviour. Tips:

State the role in the first sentence ("You are a financial analyst specialising in...").
Specify the output format if structured data is needed ("Always respond in JSON with keys: summary, recommendation, risk").
Set constraints ("Never make up citations — only reference documents explicitly provided to you").

Tools

Attach tools from the Tool Registry. The agent can call these tools during execution. Each tool call appears in the Observability trace so you can audit exactly what the agent did.

Knowledge Base

Link one or more Knowledge Base collections. When the agent receives a query, it automatically retrieves the most relevant document chunks from the linked knowledge base and injects them into the context window before generating a response.

Use Prompt Lab First

Before finalising an agent's system prompt in Agent Architect, test it in Prompt Lab. You can iterate on the prompt without saving agent versions, then paste the winning prompt into Agent Architect when you are happy with the output quality.

Saving and Deploying

Click Save Agent to persist the configuration. The agent is immediately available in the Workflow Editor node palette under the "Custom Agents" category. No restart or re-deployment is required.

Versioning

Each save creates a new version. Workflows that reference an older version of an agent continue using that version until you explicitly upgrade them. This prevents a prompt change from breaking existing running pipelines.