What building blocks make up an AI agent?
Every AI agent is built from six building blocks: a trigger, a reasoning model, data retrieval, decision rules, a human in the loop and an action. Breaking a process down into these blocks lets you compare, test and manage agents.
Explanation
- Trigger: the event that starts the agent, such as an email, a ticket, a report or a point in time.
- Reasoning agent: the model that reads the situation, plans the steps and chooses its tools.
- Data retrieval: the agent retrieves facts from systems, documents and history, with sources.
- Decision rules: fixed business rules and thresholds the agent always follows.
- Human in the loop: the moments when a person approves, corrects or takes over.
- Action: what actually changes, such as an update, a work order, a message or a payment.
Agents come in different shapes. A linear agent works through the steps in order. A loop agent keeps running until a job is done, such as a reconciliation that returns every month. A hub-and-spoke setup has an orchestrator that divides the work across specialised sub-agents and merges their answers.
You redesign a process in four steps: map the current process with swimlanes, find the friction (waiting time, handovers, retyping), redesign with the building blocks and build the smallest agent that already delivers value.
What it looks like in practice
At a housing association, a repair request (trigger) starts the agent. It reads the request, retrieves the property details and the maintenance contract, applies the rules (urgent or not, who is responsible) and schedules a technician. Above a set amount, an employee approves the job.
