Simple Agent Designer Documentation
Index

Designing AI Agents: Overview

Agents are the core building blocks of Simple Agent Designer. Each agent is a specialized AI-powered worker that can understand instructions, process data, use tools, and cooperate with other agents to complete business tasks.

What is an Agent?

In Simple Agent Designer, an agent (also called a Worker) is a configured AI entity with:

  • An identity — a unique Agent ID and title that define who the agent is
  • A specialization — a worker type that determines what kind of operations the agent can perform
  • Instructions — custom text that tells the agent how to behave and what to do
  • An AI brain — a connection to an AI provider and model that powers the agent's reasoning
  • Optional tools — the ability to call functions that interact with Business Central data

When a workload is processed, the orchestration engine sends the task to the assigned agent. The agent uses its AI provider to understand the task, follows its instructions, uses tools when needed, and returns results.

Agent Architecture

Agents in Simple Agent Designer are built on an interface-based worker pattern:

Workload → Workload Type → Initial Worker (Agent)
                              ↓
                        Worker Interface
                              ↓
                     Worker Implementation
                    (based on Worker Type)
                              ↓
                       AI Provider Call
                              ↓
                      Response + Tools

The worker interface defines three core operations:

  1. Init — Initialize the agent with its configuration, orchestration context, and workload
  2. Ask — Process an input message, call the AI, handle tool calls, and queue follow-up messages
  3. AddFile — Attach a file from the workload for the AI to analyze

Each worker type has its own implementation of this interface, optimized for its specific purpose.

The Agent List

You manage agents through the AI Agents page (also called the Worker List):

Column Description
Agent ID Unique identifier for the agent (Code, up to 20 characters)
Type The worker type that determines the agent's specialization
Title A human-readable name describing the agent's role
AI Provider Which AI service this agent uses
LLM Model The specific AI model (e.g., gpt-5-mini, claude-haiku-4-5)
Use Tools Whether this agent can call Business Central tools
General Agent Whether this is the default agent for the interactive chat

How Agents Process Work

When a workload is sent to an agent, the following happens:

1. Initialization

The agent loads its context:

  • Company information (name, country, currency)
  • Current date and GL setup
  • Agent identity (name and job title from system prompt)
  • AI setup instructions with company context
  • Agent-specific custom instructions
  • Workload type instructions
  • The workload description (as the initial user message)
  • Data context (if the workload targets a specific record)

2. AI Communication

The agent sends the assembled conversation to its AI provider:

  • System messages with identity, instructions, and context
  • User messages with the workload task
  • Previous conversation history (for multi-turn interactions)

3. Response Processing

When the AI responds:

  • If tool calls are included, the agent executes the tools and sends results back to the AI (this loops until no more tool calls)
  • If the response contains a delegation JSON block targeting another agent, the message is queued for that agent via the orchestration engine
  • The response is logged in the process activity record

4. Worker Cooperation

To hand work to another agent, the AI emits a fenced json code block containing an object with a to field (the target Agent ID) and a payload` object, for example:

{ "to": "WORKER2", "payload": { "task": "Look up customer balance for CUST001" } }

The orchestration engine parses fenced blocks in the response, reads the to and payload` values, and then:

  • Queues the payload as a message for the target worker via the ask queue
  • Initializes or resumes the target worker
  • Sends the message via that worker's Ask procedure
  • Returns results back to the original agent

If the to value does not match a configured agent, the engine reports an error back into the conversation.

One Agent per Role

A best practice in Simple Agent Designer is to create focused agents with specific roles:

  • Analyzer Agent — reads and understands incoming data
  • Data Lookup Agent — queries Business Central for reference data
  • Entry Creator Agent — creates journal entries or documents
  • Reviewer Agent — validates results before completion

This separation of concerns makes your workflows more predictable and easier to debug.