Design bounded workflows and agent systems with explicit state, tool permissions, recovery, and evaluation. Compare a fixed workflow with model-directed actions
Last reviewed: 2026-10-03
Agent architecture defines how models, tools, state, and review steps work together. A workflow follows predefined paths; an agent can choose its next action dynamically. Begin with a single bounded workflow and add multiple agents only when a measured task benefits from that coordination.
For agentic workflow design, specify tool permissions, stop conditions, recovery, and evidence of completion. For a proposed R&D exercise, use permitted research documents to build a cited literature-triage assistant with human review. That exercise demonstrates information handling, not autonomous scientific discovery or permission to conduct experiments.
When specialization, parallelism, or review structure pays: distinct skills per role, independent subtasks, or checker agents catching worker errors. Single capable agents win on simple flows: multi-agent adds coordination cost.
Reliability under autonomy: error recovery, state consistency, and bounding cascade failures. The model is rarely the bottleneck; the systems engineering around it is.