LLM Agents Through the Eyes of a Ten-Year Automator

I've been writing automation for ten years. Thousands of lines of workflows in virtualization orchestration engines, provisioning pipelines, the APIs behind self-service portals. Now everyone is saying the same thing: "LLM agents will solve everything — writing workflows is history."

Part of that is true. Part of it sounds exactly like the people who said "automation will finish every job" back in 2016.

These are field notes from someone who has seen both worlds from the inside: what changed, and what stays the same?

What changed: we describe intent, not steps

A classic orchestration workflow is a contract: inputs defined, steps ordered, every error branch stitched by hand. It's expensive to write, and the smallest scenario change demands maintenance. That was the biggest pain of the decade: workflows are brittle because they describe the world step by step.

LLM agents flip the equation. We no longer describe steps — we describe intent: "This VM's disk is filling up; find the root cause and propose a fix." The agent reads logs, pulls metrics, forms hypotheses. Flexibility that used to take a three-week workflow now fits in a system prompt.

That is a real revolution — no hype there.

What stays: the automator's laws

Here's the part the revolution likes to ignore: production environments reward predictability, not flexibility. The rules I learned over ten years, still binding in the LLM era:

1. Idempotency is non-negotiable. Automation that runs twice must not create two VMs. When an agent decides to "just try again," the thing that guarantees safety is the tool's design — not the model's intelligence.

2. Blast radius is bounded by tooling, not by prompts. Writing "don't touch production" into a prompt is not security. Whatever the agent's identity is allowed to delete, the agent will one day delete. IAM bounds authority; polite sentences don't.

3. Dry-run and human approval are not dead. For critical changes, plan → approve → apply is as true for agents as it is for Terraform. The difference: we can now read the plan in natural language. That doesn't make the approval step obsolete — it makes it more valuable.

4. Automation without an audit trail is debt. If you can't answer "why did the agent make this call?", you can't sell that system to an enterprise. Decision logs, tool-call records, rollback plans — all of it matters more than before.

The architecture that works: flexible brain, deterministic hands

The pattern that holds up in the field: the LLM resolves intent and plans; execution belongs to parameterized, testable, idempotent tools. What we call tool-calling is our old friend the workflow in new packaging — every tool has a schema for input, verifiable output, and bounded authority.

Don't put the agent in the orchestrator's seat; put it in front of the orchestrator. A translator that takes natural language in and emits deterministic steps out. Then a hallucination costs you a rejected plan — not an incident.

Closing

In 2016, automation looked like magic too; then discipline arrived, and that's when the real value showed up. LLM agents are walking the same road, just faster. The people investing in discipline rather than magic will finish this lap ahead as well.

I build tools on these principles — some of them will appear on this site soon. Let the work speak.