Opinion: LLM-first vs code-first debate in production AI
An author argues that the debate over LLM‑first versus code‑first architectures is misplaced. In an LLM‑first system the model sits at the center, deciding which tools to call and in what order. In a code‑first system, software remains in charge of sequencing, validation, permissions, and execution, while the model is only called for language understanding or generation.
Key points
- LLM‑first places the model at the center, deciding tool use and order.
- Code‑first keeps software in charge of sequencing, validation, and execution.
- Hybrid architecture uses LLM for interpretation and code for enforcement.
The piece proposes a hybrid pattern. The LLM handles interpretation, reasoning, classification, and proposal, while deterministic code validates, authorizes, calculates, persists, enforces, and executes. The author cites Google Cloud’s agentic AI guidance, Anthropic’s Building Effective AI Agents guidance, and Microsoft’s agent framework as examples of how to pair dynamic AI behavior with strict controls. He also notes OWASP’s generative‑AI security guidance, which warns against giving models excessive autonomy.
Ultimately the author suggests an intent‑first architecture, where the system first understands the user’s goal with an LLM, then converts that fuzziness into structured, deterministic operations. He stresses that reliability is an architectural property, not a prompt‑engineering trick, and that production AI should place intelligence at the right boundary.
LLM-First or Code-First? Where Intelligence Belongs in Production AI
Unite.AI · 24 September 2026
Loading the full article…
This text was published by Unite.AI. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Enterprise & Industry
All →- Gemini introduces 3.8 Live with Live Avatar · 2 src
- Oracle sends force majeure notice for New Mexico Stargate data center · 1 src
- BNP Paribas signs five-year agentic AI deal with Google Cloud · 1 src
- Anthropic partners with OpenEvidence to bring free medical AI to 100 countries · 2 src
- Trebellar raises $18M Series A to add AI agents to corporate real estate · 1 src
Comments
via GitHub Discussions