AI accelerates. Humans still decide.
Notes on why every agent I ship has a human-in-the-loop boundary — and why the cost calculus for autonomous AI is often worse than it looks.
AI can dramatically speed up QA, development, and operations. Anyone who's worked with the recent generation of models can feel it. But that speed surfaces something easy to miss:
AI hallucinates. And when left fully autonomous, it can confidently optimize the wrong thing, create unnecessary complexity, or even introduce new bugs while trying to "improve" the system.
That's why Human-in-the-Loop review is critical.
What AI is good at
- Accelerating repetitive work
- Generating ideas and drafts
- Scaling execution
What humans are still on the hook for
- Prioritization
- Business context
- Risk understanding
- Final decision-making
- Knowing what actually matters to the user
The cost trap
Sometimes the compute / token cost of uncontrolled AI workflows exceeds the value they generate. In some cases, the operational burn becomes more expensive than the manual effort it was meant to replace.
I've seen this firsthand — an agent that scored best on the accuracy rubric was costing 3x the runner-up because it kept loop-retrying on edge cases. Without runtime cost logging, that would've shipped to prod and quietly burned the budget for months.
The Pradeep tweet that landed
Pradeep Shinde recently wrote a sharp satirical piece about agents writing test cases, reviewing PRs, attending standups, and rejecting each other's builds. In his version:
"Yesterday one agent raised a defect against another agent for not following coding standards. By evening, three more agents joined the RCA call and created a Jira epic called 'Strategic Alignment for Autonomous Failure Prevention.' Meanwhile the actual bug is still in production."
It's a joke. But it's close enough to reality to be uncomfortable.
The pattern that works
Every agent I've shipped follows the same rule: the agent assists, the human decides.
- The multimodal triage pipeline adds a Jira comment. It never closes tickets.
- The Slack RCA bot summarizes Sentry + Victoria + GitHub. It never reverts a commit.
- The AIO Test Case Creator normalizes a draft. The QA author approves before publish.
- The Codex browser rollout defaults to "always ask" for new domains.
The future is not "AI replacing humans"
It's humans building controlled, reviewed, and responsible AI systems. Balance is key.
// reposted on linkedin · february 2026 · linkedin.com/in/mrnewdelhi