Agent Graph

What is an agent graph?

An agent graph represents an agentic process as connected nodes and edges. Nodes perform work—such as running an agent, calling a tool, checking a condition, or requesting approval—while edges define which transition can happen next.

The graph makes branching, loops, dependencies, and parallel paths explicit.

How an agent graph works

A support graph might contain nodes for intake, order lookup, policy retrieval, resolution drafting, approval, ticket update, and completion. Conditional edges can route high-value refunds to human approval and ordinary cases directly to execution.

Graph state carries the information needed across nodes. Each node reads selected state, produces an update, and chooses or triggers the next transition.

Agent graph versus workflow

An [agent workflow](/glossary/agent-workflow/) is the broader structured process through which work is completed. A graph is one way to represent that workflow.

Some workflows are simple sequences and do not need a graph model. Graphs are useful when execution has branches, loops, parallel tasks, or several terminal states.

Agent graph versus agent loop

The agent loop describes the repeating decision cycle of reason, act, observe, and update. An agent graph describes the possible execution topology.

One graph node may contain an agent loop, or graph edges may represent the loop's repeated transitions. The concepts operate at different levels.

Graph design concerns

Nodes need clear inputs, outputs, side effects, and retry behavior. Edges should define allowed transitions and stopping conditions. Write operations require duplicate protection when a node can retry.

An overly detailed graph becomes difficult to maintain. A graph that hides too much behavior inside large agent nodes becomes difficult to inspect. Builders should place stable rules in explicit transitions and reserve agent judgment for genuinely ambiguous decisions.

Why it matters

Agent graphs make complex execution visible and testable. They help systems track state, recover from failures, and explain why a particular path occurred.

They do not automatically create reliability. Good graph design still needs permissions, state validation, observability, and clear ownership. See [Single-Agent vs Multi-Agent Systems](/single-agent-vs-multi-agent-systems/) for how graph execution can coordinate one or several agents.