Insight 11 Mar 2026
Part 2: Human-in-the-Loop Ops: How to Get Most of the Benefit of AI Agents Without Autopilot Risk
Unattended execution is where operational risk spikes. This article outlines a practical Human-in-the-Loop workflow that captures most of the speed and clarity benefits of AI while keeping control, verification, and accountability where they belong: with the operator.
ayonik engineering
If Part 1 was about why autopilot can go wrong in production, Part 2 is about what to do instead.
The goal isn’t “no automation.” The goal is: AI accelerates thinking and drafting; humans approve execution. That delivers most of the value - faster diagnosis, better plans, fewer missed steps without turning infrastructure changes into an opaque chain reaction.
The HITL workflow for safe execution
A safe, repeatable pattern to let AI execute commands looks like this:
- Describe the outcome (what “good” looks like)
- Collect facts (read-only commands first)
- Propose a plan (steps + reasoning)
- Prepare rollback scenarios
- Preflight / verify (what could break? what will change?)
- Execute (only after explicit confirmation)
- Validate (prove it worked)
- Document (what happened, why, how to recover)
This is boring on purpose. It matches how good ops already works - AI just helps compress the thinking and writing time.
Default to read-only first
A practical guardrail: AI never starts with write actions.
Good “first move” categories (depending on the intent):
- environment identification
- service status and recent logs
- process/socket state
- disk/memory/CPU snapshots
- recent system events
- dependency checks (DNS/TLS/upstreams)
Preflight checks: “What will change?” before “Change it.”
Preflight is how you keep velocity without losing safety.
Common preflight patterns:
- config syntax check before reload/restart
- dry-run modes where available
- diff/preview before apply
- scope confirmation (hostnames, clusters, environments)
A good plan explicitly calls out:
- which files will be modified
- which services will restart
- expected downtime impact
- how to revert (or what backup exists)
Approval boundaries: approve steps, not vibes
Human confirmation only makes sense when boundaries are clear—meaning exact commands and scope are visible before execution.
Confirm a scoped action, not an intent.
Batch safely: canary → small batch → full rollout
If changes must happen across multiple machines, a rollout ladder reduces risk:
- Canary host
- Small batch (2–5)
- Full rollout
- Post-checks everywhere
The AI tool should produce:
- a rollout plan
- a rollback scenario
- a validation command set
- stop conditions (“if X happens, stop and rollback”)
Documentation as an output
A quiet superpower of Human-in-the-Loop AI is turning ad-hoc fixes into something repeatable.
After validation, the tool can produce one of:
- a short incident wrap-up
- a runbook section (“Symptoms → Checks → Fix → Verify → Rollback”)
- a postmortem draft with concrete commands and findings
- a script to answer to similar issues in the future
This is how you compound value over time.
Implementation note
The hard part isn’t knowing the workflow - it’s following it consistently under pressure. Admin Companion helps by structuring suggestions (plan, checks, risks, validation) and requiring explicit confirmation before execution, so the review step can’t be skipped accidentally.
https://www.admin-companion.ai/product
Next in this series:
Part 3: Human-in-the-Loop AI Agents: A NIST-Based Risk View
First article in this series
AI agents are increasingly positioned as the next step for sysadmins: automate routine requests, triage alerts, apply changes, and reduce operational toil. Parts of this promise are real, but there is a gap between agent demos and production reality. This article explains where autonomy helps, where “autopilot” becomes risky, and why Human-in-the-Loop often delivers most of the benefit without surrendering control.
Human-in-the-Loop is not a “less advanced” form of AI agents - it is a deliberate risk and accountability design choice. Using NIST’s AI Risk Management Framework and its Generative AI Profile as a lens, this article summarizes what trustworthy AI requires and maps those expectations to a terminal-native HITL assistant like Admin Companion - highlighting both strengths and intentional boundaries.