Agent Placement, Not the Model, Decides If You Get an Agent or a Chatbot
bammcd_builds · reddit · 2026-08-13
Developer Brian McDonald explored the impact of architectural choices on agent capabilities by building the same AI agent three different ways on the ServiceNow platform: external over MCP, interactive via desktop client, and natively inside the platform.
The experiment revealed that while the underlying model remained identical, the agent's placement and identity determined whether it acted as a true agent or just a chatbot. The native agent inherited the platform's own permission model (ACLs) without needing a custom authorization layer, allowing it to perform real actions (like placing service catalog orders) directly via API, at the cost of portability. Conversely, the external MCP approach was portable but required managing independent authorization logic.
The author concludes that while most agent architecture discussions focus on models or frameworks, the actual differentiator between a production-ready agent and a demo-only bot is where it runs and under what identity.
More from coding & agent
- AI Digital Coworkers Emerge as New Meta: OpenAI, Grok and Others Join In — tushaarmehtaa · 2026-08-13
- Grok 4.6 in Action: 10 Wild Coding and 3D Generation Demos Within 48 Hours — minchoi · 2026-08-13
- Microsoft Paper Reveals Hidden Costs of Bad Skills in AI Agent Harnesses — omarsar0 · 2026-08-13
- Stop Waiting for Better Models: Architecture is the Key to Agent Reliability — import_jmr · 2026-08-13
- NVIDIA Offers 7 AI Courses for Free, Covering Agentic AI and Robotics — ifioknkem · 2026-08-13
- Claude Code's harness makes almost everything worse, says user — aiamblichus · 2026-08-13