Autonomy, tool use, memory and multi-step reasoning make an agent capable. The agentic loop is what makes it operational. This lesson shows how those properties combine into a repeating cycle of action and evaluation that allows agents to handle real work, not just respond to prompts.
Traditional software works in a straight line. You give it an input, it produces an output, and it stops. If something unexpected happens mid-process, it waits for you.
Agents work differently. They work in a loop. They take a step, evaluate what happened, and decide what to do next. They keep going until the task is done, or until a point is reached where a human needs to make a call.
This loop is not just a technical detail. It is the mechanism that allows agents to handle real-world complexity. Understanding is the key to understanding what agents can and cannot do.
Stages of the Agentic Loop

The Key Idea: The loop enables an agent to handle real-world complexity. Rather than following a fixed script, the agent adapts at each step based on what it finds.
The loop has six stages:
- Perceive - The agent gathers what it needs to understand the task — reading relevant data, checking the current state of systems, and understanding any context from prior steps.
- Plan - The agent maps out how it will approach the task: what steps to take, in what order, and which tools to use. This plan can adjust as new information comes in.
- Act - The agent takes one action — querying a system, updating a record, triggering a workflow. It only does what it is authorized to do, and every action is logged.
- Observe - The agent checks what happened. Did the action succeed? What changed? This is how it builds an evidence trail and knows what to do next.
- Reflect - The agent asks: am I on track? Does this align with the goal and the rules? If something is off, this is where it catches it before moving forward.
- Repeat - If the task is not finished, the loop runs again from the top. When it ends, the task is either complete and documented, or handed to a human with full context already loaded.
Human Touchpoints in the Loop
The agentic loop does not mean the agent runs indefinitely without human involvement. Organizations define precisely when the loop pauses:
- When a decision exceeds a defined financial threshold.
- When a customer-facing action requires review.
- When the agent encounters an exception outside its policy boundary.
- When a regulatory requirement mandates explicit approval.
At each of these points, the agent stops, surfaces the relevant context, and waits for a human decision before proceeding. This is not the agent's fault. It is governance by design.
Here is how the loop operates in a real enterprise scenario. An invoice arrives flagged for a price discrepancy against the original purchase order.
Illustrative Walk-through - Invoice Discrepancy

With this topic, track which stage of the loop each action maps. Notice where the agent adapts instead of stops.
Perceive: The agent reads the invoice, retrieves the purchase order from the ERP, and checks the goods receipt record.
Plan: It maps the steps: compare line items, check tolerance thresholds, and identify the source of the discrepancy.
Act: It cross-references the invoice amount against the purchase order terms and contract pricing.
Observe: The discrepancy is 2.3%. The tolerance policy allows up to 3% for auto-resolution.
Reflect: The discrepancy falls within policy. No escalation required. Proceed to resolution.
Act (again): The agent posts the matched amount, logs the resolution, and closes the invoice in the ERP.
Resolve: Task complete. Full audit trail generated. Finance team notified.
If the discrepancy had been 4.7% — above the tolerance threshold — the agent would have paused after Reflect, surfaced the full context to a finance reviewer, and waited for explicit approval before proceeding. The loop handles both paths.
When Agents Work Together
Some enterprise processes are too large and too cross-functional for a single agent to handle effectively. For these, multiple specialized agents coordinate under a central orchestrating agent.
Think of it as a well-run project team. One coordinator receives the overall goal, breaks it into parts, assigns each part to the right specialist, and tracks progress until the whole task is complete. Each specialist only needs to know their piece.
Joule is designed to be exactly this orchestrating layer. It coordinates specialized agents across Finance, HR, Supply Chain, Procurement, and Customer Experience, with each agent grounded in the relevant SAP process context and data. This is not a future capability. It is how Joule Agents operate today.
Lesson Summary
- Agents work in a loop: perceive, plan, act, observe, reflect, repeat. Not in a straight line.
- The loop is what allows agents to handle real-world complexity: unexpected results, missing data, exceptions.
- Human touchpoints are built into the loop. The agent pauses when the organization requires it.
- Multiple agents can coordinate: one orchestrator, multiple specialists, for complex cross-functional work.
- Joule is designed to orchestrate these agent teams across the entire SAP portfolio.
Note
Unit Summary
In this unit, you examined the foundational properties and workflows that allow agents to operate effectively. You learned that autonomy, tool use, memory, and multi-step reasoning work together as an integrated system to enable agents to perform complex work.
You also explored the agentic loop and how agents continuously perceive, plan, act, observe, reflect, and repeat as they work toward a goal. Finally, you examined how multiple agents can coordinate under an orchestrator to tackle cross-functional business challenges.
You should now be able to:
• Describe the four core properties of an AI agent.
• Explain the stages of the agentic loop and how agents execute tasks.
• Understand how orchestrated agents collaborate to solve complex enterprise problems.