I keep seeing the same demo.
An engineer, proud. "Watch this," they say. "The agent takes the Jira ticket, writes the code, creates the PR, and deploys to staging. Fully automated."
It works. Impressively fast.
But nobody ever demos what happens when it's wrong.
The Pattern I Keep Seeing
Every organisation racing to adopt AI agents makes the same mistake. They treat agents as magic.
"The agent handles it."
"We automated that with AI."
"It just works."
Until you're pulled into that incident channel.
What Leaders Forget
An AI agent is operationally equivalent to your fastest, most reckless engineer.
The one who ships in hours what takes others days. Who knows where the bodies are buried. Whose colleagues can't keep up.
Now imagine that engineer had no PR process. No code review. Direct production access. No approval workflows. No audit trail.
That's what ungoverned AI agents are.
The Governance Principle
Here's what most teams miss: AI agents need stricter controls than humans. Not the same. Stricter.
It feels counterintuitive. If the agent follows the same PR process, the same review cycle, the same deployment pipeline - surely that's enough?
It's not. Because humans come with built-in friction. We get tired. We second-guess ourselves. We ask colleagues before doing something irreversible.
Agents don't have that friction. So you have to build it externally.
| Process | Human Developer | AI Agent |
|---|---|---|
| Pull Requests | Required | Required |
| Code Review | Peer review | Mandatory expert review |
| Elevated Access | Manager approval | Explicit scope + expiry |
| Production Deploy | Standard pipeline | Same pipeline, additional validation |
| Audit Trail | Git history | Git + agent decision logs |
| Rollback | Standard | Automated + faster |
The faster the agent, the more critical the guardrails.
What This Looks Like in Practice
Pull Requests: No Direct Commits
I don't let agents push directly to main. Every change goes through a PR. I haven't found a good reason to make exceptions. The PR is your audit trail. It's where humans verify before code hits production.
I've seen teams give agents commit access "to move faster." They moved fast alright. Straight into a production incident.
Code Review: Humans Approve, Always
AI-assisted review can triage. Flag obvious issues. Check for patterns. But a human still approves. Every time.
Why? Because an agent that writes code and approves its own code has no external check. It's confident in its own work. That confidence means nothing.
Scoped Access: Least Privilege, Hard Expiry
I don't give agents standing access to critical components. They get scoped permissions for specific tasks. Those permissions expire.
An agent that "needs" persistent access to critical systems is an agent whose blast radius you can't control.
Decision Logs: What Did It Do and Why?
Git history shows what changed. Agent decision logs show why.
When something breaks at 2am, you need to know: what did the agent decide? What context did it have? What did it miss?
Without decision logs, you're debugging a black box.
Rollback: Faster Than Deployment
If an agent can deploy in minutes, you need to roll back in seconds.
I wire every agent deployment with an automatic rollback trigger. Error rate spikes? Rollback. Latency jumps? Rollback. Don't wait for a human to notice.
Why This Matters Now
When a human makes a mistake, they usually notice something felt wrong. They hesitate. They ask a colleague. The social friction of development creates natural checkpoints.
AI agents don't hesitate. They don't feel uncertain. They execute with confidence whether they're right or wrong.
That confidence will ship bugs just as fast as features. Your process is the only check.
The Uncomfortable Truth
Organisations that skipped governance for human developers will find AI agents unforgiving.
The shortcuts you tolerated before become catastrophic at AI velocity.
Questions to Ask Yourself
Before you deploy an agent:
- Would you give a new contractor this level of access?
- This much autonomy?
- With this little oversight?
If the answer is no, your agent governance isn't ready.
Thoughts? I'm on X.