When people describe an AI system, they often draw either a cycle or a network. The cycle shows how work repeats. The network shows how pieces depend on one another. These are not competing diagrams. They are two different engineering lenses.
The difference in one sentence.
Loop engineering designs the repeated process through which a system observes, decides, acts and learns. Graph engineering designs the structure of entities, dependencies and possible routes through which work can move.
Optimises repetition, feedback and improvement over time.
Optimises relationships, dependencies and branching choices.
Loop engineering creates momentum.
A loop is the operating rhythm of a system. A support agent receives a question, gathers context, proposes an answer, checks the result and learns from the outcome. A product team observes behaviour, forms a hypothesis, ships a change and measures what happened.
The important engineering question is not simply whether each step works. It is whether the output of one cycle makes the next cycle better.
A well-designed loop defines:
- Entry conditions: what event starts the cycle.
- State: what the system must remember between steps.
- Feedback: what evidence tells the system whether it succeeded.
- Stopping rules: when to finish, escalate or ask for help.
Key ideaLoops create improvement only when feedback changes the next decision.
Graph engineering creates structure.
A graph represents things as nodes and their relationships as edges. In a software project, nodes might be requirements, components, data sources, tools and tests. The edges reveal which pieces rely on others and which routes are available.
This becomes valuable when a workflow is not truly linear. Research can branch into several hypotheses. An agent might select different tools depending on the evidence. A failure in one component can affect multiple downstream actions.
A well-designed graph makes visible:
- Dependencies: what must exist before another step can begin.
- Branching: where evidence can send work down different routes.
- Ownership: which person, agent or tool controls each node.
- Impact: what else changes when one node changes.
The strongest systems use both.
A graph without a loop can describe a sophisticated system that never improves. A loop without a graph can repeat quickly while hiding fragile dependencies and poor choices.
Imagine an AI research agent. Its graph defines sources, claims, evidence, tools and relationships between findings. Its loop defines how it searches, evaluates evidence, identifies gaps and repeats until the stopping conditions are met.
gives the system a map.
Loopgives the system a pulse.
A practical design sequence.
When building an AI-enabled product, begin with the graph because it exposes what exists. Then design the loop because it explains how the system moves and improves.
- 01Name the nodes.
List the users, data, decisions, tools, outputs and checks involved.
- 02Draw the relationships.
Connect dependencies, ownership and possible branches.
- 03Trace one complete loop.
Follow a real task from trigger through action, verification and learning.
- 04Define failure paths.
Decide what happens when evidence is missing, confidence is low or a tool fails.
- 05Measure the next cycle.
Choose feedback that can meaningfully improve the following run.