Adding Edges Is Not a Paradigm Shift

The post in question highlighted that adding edges is not a paradigm shift in the field of agent chains. After initially agreeing with the post and planning a migration, the author found that the sequence they had been working on already had a fixed structure with checkpoints between each step and only one edge leaving every node. The author realized that this structure was not a loop, but rather a directed cyclic graph, which is the basic form of a graph.

The post went on to explain that the hype around the term "graph engineering" was due to a shift in vocabulary, rather than a fundamental change in the underlying design. The author pointed out that production agents require cycles, as they retry failed calls, ask for missing information, correct themselves, and pause for human input. Thus, a loop is a necessary component for these agents, making it the "simple case" of a graph.

The author further argued that loop engineering is not an alternative to graph engineering, but rather the basic form of it. They cited examples of prior art, such as Anthropic's documentation on prompt chaining and Microsoft's development of graph workflows, which were all published before the term "graph engineering" gained popularity.

The author concluded by emphasizing that the real question to ask is not whether to adopt graphs, but rather, what each additional edge will cost in terms of complexity and overhead. They provided an example of a seven-step agent chain, where each step has a checkpoint where the output is read by the next step. Adding additional edges and branches would result in multiple nodes building on the same spec in parallel, leading to potential errors surfacing at convergence rather than checkpoints.

The author emphasized that gating is not free and becomes expensive when multiple branches are involved, making a simple chain a more cost-effective solution.