Flowcharts vs. Handoffs: a simple math framing (blog.rowboatlabs.com)

The post contrasts two architectures for orchestrating multi‑agent AI systems: visual flowchart builders (n8n, OpenAI’s Agent Builder) where control moves along a statically drawn graph, versus a handoff model where any agent can pass control to any other and the entire conversation history H travels with that handoff. The authors argue that handoffs produce a compact, dynamically generated call graph that grows with the task, making open‑ended delegation, retries, and escalation rules simple functions over H. They illustrate this with a Searcher→Responder→Verifier example in which agents choose next steps at runtime based on coverage and confidence; adding new agents or failure paths in a flowchart quickly creates combinatorial edge/guard blowups, whereas in the handoff model you usually add only an agent function and a small routing rule. They also note they implemented Rowboat on OpenAI’s Agents SDK and open‑sourced it. Technically, the flowchart is a fixed directed graph G=(V,E) with branching decided at design time, while the handoff model defines a finite agent set A and treats each agent as a function fa: H → (H', next_agent ∈ A ∪ {halt}), i.e., continuation‑passing with state. Consequences: history is first‑class in handoffs (no fragile threading of variables), routing logic is local to agents (dynamic routing without enumerating n*(n−1) edges), and description cost is O(n) versus often O(n^2) for flowcharts as tasks become open‑ended. Bottom line: flowcharts suit fixed pipelines; handoffs provide greater expressiveness, maintainability, and extensibility for open‑ended, adaptive workflows.