Agentic Workflows Should Mature Into Deterministic Code

New research on "Progressive Crystallization" describes a lifecycle where agentic workflows explore novel problems freely, then get demoted into deterministic code once traces prove repeatable behavior. In a cloud-network operations case study, deterministic execution rose from 0% to 45% over eight months, cutting per-incident agent cost by more than 70% while incident volume doubled.

The economic logic mirrors the Compiled AI paper's function-calling results: once a workflow is compiled, it reaches 96% task completion with zero execution tokens, breaks even at 17 transactions, and cuts token usage 57x at 1,000 transactions. Agentic runtime inference becomes a per-run tax; once a workflow repeats often enough, compiling it into deterministic code removes that tax.

LangGraph's Functional API (@entrypoint, @task) is presented as the practical seam between agent exploration and compiled execution, supporting checkpointing, idempotent writes, and typed failure states without forcing a binary choice between free-form loops and rigid workflow engines.

The paper stresses that promoting a workflow must rest on evidence — a corpus of traces plus passing acceptance tests — not on a model's own claim that a path is repeatable.