Single-User vs Multi-User Agents: What Actually Changes

Building an AI agent demo is very different from deploying it to many users. Single-user agents can rely on simple state (markdown files), tolerate inefficiency, and skip guard rails. Multi-user agents require state isolation to prevent context leakage between users, cost controls and quotas to prevent runaway spending, latency management with queues and retries, observability and audit logging, and security measures against prompt injection and tool misuse. The core distinction is framed as 'agent core' (planning, memory, tools) vs. 'agent harness' (isolation, auth, cost controls, observability) — single-user systems are core-heavy while multi-user systems must nail the harness first.