Architecture
Models are only one layer. A walk through the request path, and the shared services that keep agents reliable.
Oct 29, 2026 · 6 min read
Demos fail at the same place: the step from a convincing prototype to a dependable system. The gap is rarely the model. It is the layers around it: how work is orchestrated, what the agent knows, who it acts as, and how you see what happened.
Enterprise agent architectures usually separate models, orchestration and communication protocols, knowledge and data, and governance and observability. Each can change on its own, which matters because models change every few months.
A message arrives from a channel such as WhatsApp or a phone line. The gateway authenticates it and applies limits. The orchestrator breaks the task into subtasks and delegates to specialist agents, calling models through the router. Specialists reach business systems only through the tool layer, where connectors are scoped to least privilege. Every step is traced and results are scored against evaluations.
The AgentArch benchmark, published by ServiceNow researchers in 2025, found that even strong models reached only 35.3% success on complex enterprise tasks, and that choices about tools and memory changed results materially. The lesson is to design per task and measure, instead of copying one agent template everywhere.
Cost and risk matter too. Gartner has predicted that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. Governance, measurement and a business metric belong at the start.
Tell us about one process. We will show how an agent could handle it, and what to measure to know it works.