love your llm, don't abuse your llm
CI/CD has been solved. Not perfectly, but well enough. You push code, tests run, artifacts build, things deploy. The pipeline is boring and that’s the point. Boring infrastructure ships product.
Then LLMs arrived and everyone started asking: what if the LLM ran the pipeline?
Wrong question.
Abuse is making something do what it was never meant to do. An LLM is not a task runner. It is not a build system. It is not a deployment orchestrator. It is a reasoning engine — extraordinarily capable when reasoning is what’s needed, wasteful and brittle when you use it to replace a cron job.
Your CI/CD pipeline doesn’t need to become an LLM. It needs to grow up.
Grow it. Add stages. Improve observability. Reduce feedback loops. Let it be what it is: a deterministic, auditable, reproducible conveyor belt.
Then wire your Agents into it.
Agents (capital A — autonomous, goal-directed, tool-using) belong inside the pipeline as participants, not as replacements for it. An Agent that reviews a diff before merge. An Agent that writes release notes from commit history. An Agent that opens a ticket when a test flake crosses a threshold. An Agent that proposes infra changes in response to cost alerts.
These are natural. These are leverage.
What’s not leverage: an LLM deciding whether to run terraform apply. What’s not leverage: prompting your way through a deployment because you never invested in the pipeline itself.
Love your LLM. Use it for what it’s good at — reasoning, synthesis, judgment under ambiguity. Build the pipeline around it. Wire the Agents in.
Don’t make the LLM do your ops homework.