Back to AI Research

AI Research

FAOP formalizes agent workflows as typed steps with scoped tools

Key Takeaways

  • The theory introduces four normal forms and sound rewrite rules, with an implementation and a worked example described in the paper.
  • FAOPs: Normal Forms and an Algebra for AI Agent Workflows proposes a formal way to organize multi-step agent procedures.
  • The author identifies recurring problems in coarse files: several actions bundled into a step, copies that drift apart, broad tool access and refactoring without checks.
  • The approach represents a workflow as a typed function over scoped state.
  • Its purpose is to expose contracts and relationships that are difficult to inspect when a procedure remains an undifferentiated block of instructions.

Make workflow structure explicit

FAOPs: Normal Forms and an Algebra for AI Agent Workflows proposes a formal way to organize multi-step agent procedures. The author identifies recurring problems in coarse files: several actions bundled into a step, copies that drift apart, broad tool access and refactoring without checks.
The approach represents a workflow as a typed function over scoped state. Its purpose is to expose contracts and relationships that are difficult to inspect when a procedure remains an undifferentiated block of instructions.

Four normal forms address different problems

The Decomposed form splits work into atomic steps. Typed makes data flow explicit and subject to type checking. Factored gives a shared procedure one canonical definition, while Scoped limits each step's tool access.
These forms are progressive structural requirements, rather than four names for the same quality check. A workflow can have small steps while still passing incompatible data or granting those steps more authority than they need. The theory attempts to make each of those concerns independently inspectable.

Refactoring has assumptions and boundaries

The paper adds an algebra of sound equalities for refactoring and parallelization, with extensions for termination, optimization and time or cost bounds. It also describes a classifier for deciding whether a workflow needs AI at all.
The author implements the approach in SkillBase with a separate executor and illustrates it through a worked financial workflow. That example is an illustration of the framework, not evidence of a measured production outcome or a general compliance guarantee. Formal properties depend on the declared contracts and assumptions being satisfied.
Franklin's reading is to begin by checking whether a real workflow exposes its inputs, outputs and permitted tools at each step. Those declarations make later testing and changes easier to reason about. They do not, by themselves, establish that a language-model answer is factually correct or that every external tool performs as specified.
Read the FAOP paper.

Comments