# Cisco MyAgent: The Enterprise AI Architecture to Copy > Source: > Published: 2026-09-04 08:16:19+00:00 Cisco gave all 90,000 of its employees a personal AI agent last week. The number is eye-catching. The architecture behind it is more interesting. The agent is called **MyAgent**. It runs on Circuit, Cisco’s internal AI platform, and does what most enterprise AI promises do not actually deliver: it acts. Employees set an objective, and MyAgent coordinates the steps needed to complete it across Outlook, Webex, Jira, and SharePoint — with no tool-switching required. But the real disclosure is buried one level down: how Cisco routes AI requests to make this economically viable at scale. ## The Cost Routing Breakdown Cisco disclosed its model routing split, and it is worth reading slowly: **50–60%** of requests go to open-weight models hosted on Cisco’s own GPU servers**20–30%** go to classic software automation — no language model involved at all- A **residual fraction** reaches external frontier APIs: Azure OpenAI, Claude, Gemini Think about that. The overwhelming majority of agent work at Cisco never touches a paid frontier API. Open-weight models running on-premises handle the execution. Software automation handles the deterministic steps. Frontier models get the complex planning work that actually requires them. This is the most important enterprise AI cost architecture story of 2026. Using a single premium model for everything is not a plan — it is a budget leak. Cisco’s numbers validate what AI cost researchers have been saying: optimized [multi-model routing cuts API costs by 60–80%](https://www.requesty.ai/blog/ai-agent-cost-optimization-how-to-cut-llm-spend-by-80-percent-with-routing) compared to routing all traffic through one premium model. Cisco is validating this at 90,000-person scale. ## 800 Subagents, Not One Giant Agent MyAgent is not a monolithic AI assistant. It is an orchestrator sitting on top of more than **800 backend subagents**, each with a narrow job. One tracks sales-forecast variances. Another sends anomaly notifications. Each one does its specific task well rather than trying to do everything passably. Personal agents stay isolated from each other — your MyAgent cannot access what a colleague’s MyAgent is doing or has learned. This is the right architecture for enterprise AI. The failure mode of most agent projects is building one agent and overloading it with context. Cisco’s approach — decompose into a catalog of narrow specialists, orchestrated by a routing layer — is cleaner, more testable, and more maintainable. When a subagent fails, you replace that subagent, not the entire system. ## The Guardrails Pattern MyAgent does not act without limits. External actions and destructive operations require **explicit human approval**. A policy server enforces the rules: it can block operations entirely and prevent Cisco data from being used to train third-party models. This is supervised autonomy done right. Agents act freely within defined bounds; they escalate anything outside those bounds. Cisco also built this on [three years of Circuit platform investment](https://blogs.cisco.com/news/from-ai-experiments-to-90-adoption-how-cisco-operationalized-ai-at-scale) — multi-model routing, RAG-as-a-service, and permissioned connectors across 4,300 internal teams — before MyAgent launched. The 8 million quarterly interactions already happening on Circuit meant the infrastructure was proven before the agent layer arrived. ## The Tension Worth Naming Cisco announced approximately 4,000 job cuts in May 2026, redirecting investment toward AI, silicon, and security. MyAgent launched in August. CFO Mark Patterson called the restructuring “not savings-driven.” That may be true in the narrow sense. But deploying 90,000 AI agents that autonomously handle tasks in the same quarter you reduce headcount by nearly 5% is a sequence that employees notice. The honest read: the cost routing model Cisco disclosed — 50–60% handled by on-prem open-weight, 20–30% by software automation — functions as a cost optimization regardless of framing. That is not a criticism. It is the correct engineering approach. But the enterprise AI story is not complete without naming the structural shift happening alongside the productivity gains. ## What Developers Should Take From This If you are building enterprise AI agents, [Cisco’s MyAgent blueprint](https://blogs.cisco.com/news/my-agent-and-the-rise-of-ambient-intelligence-ciscos-next-step-in-enterprise-ai) gives you three things to act on now: **Tier your models.** Route planning and complex reasoning to frontier models. Route execution loops and simple lookups to open-weight models via Ollama, vLLM, or an LLM gateway. Route deterministic steps to software automation. Defaulting to the best model for everything is the expensive default.**Decompose into subagents.** A single agent that knows everything is a liability. Build narrow specialists with clear interfaces. Your orchestrator routes; your subagents execute. This is also easier to debug and swap out when something breaks.**Build your control plane first.** Cisco did not build MyAgent at a hackathon — it had three years of Circuit investment first. Permissioned connectors, policy enforcement, model routing, and observability are not features to add later. They are the foundation. Ship those before your agents. At 90,000 agents, the architecture is not academic. The numbers Cisco disclosed are a [real-world blueprint](https://www.pymnts.com/news/artificial-intelligence/2026/cisco-deploys-custom-ai-agent-to-entire-90000-person-workforce) that any team scaling enterprise AI should internalize before they run into the bill that makes it obvious.