The Agent Hype vs. Reality: What Actually Works in Production

Everyone is building multi-agent systems. Agents spawning sub-agents, model routers choosing which AI handles what, orchestration frameworks with 100-tool integrations.

I tried it. Two weeks with an agent framework. Skill routing intercepted my services. Sandbox configs diverged from the repo. Session context went stale. Sub-agents could not drive tool chains. Every new capability required debugging the framework before reaching the actual code.

Then I tried something simpler. One smart agent operating reliable, deterministic machines. Services that take arguments and return results, no framework, no sub-agents, no orchestration layer deciding at runtime which code to execute.

Two hours to get it working, and it shipped.

The architecture that works in production is boring. One operator agent that understands the business context. A set of services that each do exactly one thing. The operator calls the services. The services return results. The human steers.

The machines are the product; the agent is just the operator. The intelligence is concentrated in the human who decides what to build, not distributed across a swarm that decides what to do.

Every new capability is the same pattern: a service, a configuration, a channel. No new architecture, no new framework integration. The system scales by adding machines, not by adding complexity. The only complicated thing was the two weeks I spent learning that complicated does not ship.