Agent-or-Workflow Decision Tool
Most operational problems do not need an AI agent. An agent is slower, more expensive, and non-deterministic. This decision tree tells you frankly whether your problem calls for a simple 50-line deterministic script, a cron job, or a guarded agent pipeline.
What format does your input data arrive in?
Consider the raw source data where your operational workflow starts.
Production Rulebook: When NOT to Use an AI Agent
LLMs predict next tokens probabilistically. They cannot sum floating-point numbers or calculate tax reliably. Use deterministic SQL or TypeScript for all math.
Never allow an agent to write directly to a production database. Always buffer output into an isolated staging schema with strict CHECK constraints.
If the input source provides predictable JSON or CSV headers, an agent is an expensive gimmick. Write an automated ingestion script that costs $0/month.