B6 · Custom state graph

Family: Workflows

The flow is declared as a graph of nodes, conditions and cycles, some nodes being plain code and others LLMs.

Principle

The process is described as a graph: nodes, transitions, conditions, cycles. Some nodes are plain code, others call an LLM or an agent. The graph offers resume points, human interruptions and fine-grained traceability.

The problem it solves

A branching business process mixes precise rules and judgement. A simple chain cannot express it, and a free-roaming agent does not always follow it.

When to use it

  • The process is branching and mixes rules and judgement, with steps an LLM does better than code.
  • You need resume points, interruptions for human approval and precise traceability of every transition.

When to avoid it

  • The process is linear: a prompt chain (B1) is enough and easier to read.
  • The path is unpredictable and the graph would turn into a list of special cases: let the model decide (C1).

Common pitfalls

The anti-patterns that misuse this pattern, best established first. Each sheet describes the symptoms, the public incidents and the remedies.

Tools and agent-computer interface

Loop, stopping and orchestration

Oversight, operations and governance

  • AP82 · Opaque framework — The framework hides the prompt, the context sent or the control flow: you cannot tune what you cannot see.
  • AP83 · Stale scaffolding — A breakdown, a tool or an instruction made up for an older model; it still costs and no longer adds anything.

All anti-patterns

Synonyms by vendor

  • Custom workflow (LangChain, LangGraph)
  • Custom logic (Google Cloud)
  • Graph (AWS Strands)
  • Graph-based workflows (Google ADK 2.0)
  • Workflows (Microsoft Agent Framework)

Neighbouring patterns

Sources

Content reviewed on 2026-10-02

Pattern catalogue