Use a plain script with an LLM call when the task is linear and the inputs are predictable; reach for an AI agent only when the workflow needs real tool use, memory across steps, or branching decisions. The new llm 0.37 release from Simon Willison strengthens the script path by adding gpt-6.1-sol and gpt-6.1-terra, a new llm openai endpoint add command for OpenAI-compatible APIs, and a cheaper default model (gpt-6-luna) at half the price of the previous one.
The llm CLI is a command-line and Python library that talks to many providers through one interface. Version 0.37 also updates embeddings to handle multiple content types, which matters for models like Gemini Embedding 2 that cover text, audio, image, and video.
What this changes for small-business pipelines
Most automations I ship - invoice OCR, lead triage, document summaries - are linear pipelines, not agents. A cron job plus the llm CLI often does the work of a heavy agent framework at a fraction of the cost and complexity, which is close to how I built a tender analysis pipeline that scans hundreds of pages for hidden penalties and deadlines. The new endpoint-add command also makes it trivial to point the same code at a self-hosted model when a client needs data to stay inside their perimeter.
Why it matters: Pick the smallest tool that solves the job - a scripted LLM call for linear pipelines, an agent only when branching logic earns its weight.
Source: Simon Willison