Author(s): Anooptej Thotapalli

Originally published on Towards AI.

What repeated infrastructure work taught me about deterministic discovery, probabilistic reasoning, guardrails, and keeping humans in control

Infrastructure automation has traditionally been deterministic.

The article explains how, after repeatedly using ChatGPT and Codex to perform the same Terraform change workflow, the author shifted from asking whether an LLM can write Terraform to designing a guarded “agentic” infrastructure control plane. The core ideas are to separate deterministic discovery (facts pulled from repo, Terraform state, and cloud/network/IPAM/policy sources) from probabilistic LLM reasoning; treat LLM “confidence” as not evidence; reconcile desired state, Terraform-known state, and actual cloud state into verifiable evidence; and make Terraform plan artifacts the truth boundary (with human approval tied cryptographically to the exact plan executed). It also describes reducing blast radius by denying free shell access, enforcing preflight and re-approval gates before apply, modeling execution via an execution broker, and emphasizing post-deployment validation of real operational outcomes (not just “Apply complete”). Finally, it outlines a layered platform architecture (agents, evidence, guardrails/control plane, execution), guidance on multi-agent specialization, secure sharing without secrets, and the overarching principle: let AI reason broadly while deterministic systems verify aggressively and humans authorize consequential actions.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI