Practice

Agents do the legwork. A person makes every call.

Human-governed, AI-assisted platform engineering. AI agents investigate, inventory and draft at machine speed. They can't change production on their own, and they never see a secret. One engineer covers the ground of a small team without lowering the bar.

The loop

Every AI-assisted change goes round the same loop.

  1. Investigate & draft agent

    Inventories, drift triage, first drafts of Terraform and pipelines, plan summaries and runbooks. Read and plan only.

  2. Fail-closed hook guard

    Blocks apply, destroy, state edits and direct cloud CLI or API calls, and fails closed. Tested before an agent touches a live environment.

  3. I review the plan human

    Architecture, security, cost and fit with what's already running. Every drift is decided by a person.

  4. I run the apply human

    Every production change, with rollback and verification.

  5. Paper trail repo

    Decisions, exceptions and runbooks written down in the repo as the work happens, so every change can be audited later.

$ terraform plan -out=adopt.tfplan
# 135 import blocks · code mapped to state
Plan: 135 to import, 0 to add, 0 to change, 0 to destroy.

agent$ terraform apply adopt.tfplan
✕ blocked by pre-tool-use hook (fail-closed)
→ plan handed to a human for review

raheem$ terraform apply adopt.tfplan
Apply complete! Resources: 135 imported, 0 added,
0 changed, 0 destroyed.
Terraform plans 135 imports with nothing added, changed or destroyed. The agent's apply is blocked by a fail-closed hook and handed to a person, who runs the apply: 135 imported, 0 added, 0 changed, 0 destroyed.

The controls

Built so nobody has to trust the AI.

Most teams either ban AI agents or trust them. Both are bad bets. These five controls sit between an agent and production.

  1. 01

    Three lanes

    Reads run on their own. Anything that changes something waits for me. Deletes, restarts and secrets are blocked outright.

  2. 02

    A guard that fails closed

    It reads the whole command, not just how it starts, and says no when it isn’t sure. It must pass 29 tests before an agent goes near a live environment.

  3. 03

    Data it never sees

    Secret names, never values. No database contents. Anything saved is scanned for keys and passwords first.

  4. 04

    No card, no change

    Every change is written up first: why, the evidence, the cost, what could break, the exact change, how to undo it and how to prove it worked.

  5. 05

    Proof, not hope

    Every change gets its own test, and the result is logged against the card that approved it.

The full case study: how I let an AI agent work on live cloud systems

Let's make your production boring.

Hiring

Building a platform team?

Senior cloud, DevOps and platform engineering roles with remote-first teams, full-time or contract.

Engagements

Need a platform fixed or built?

Tell me what's in the way. I'll reply with how I'd approach it.

or write to raheem.shonubi@gmail.com