The Human in the Loop
Planning autonomy levels for AI executives
"So the AI just... does things?"
I get this question a lot when I explain the vision. Usually with a hint of concern in the voice.
And the answer will be: yes and no. Mostly no. It depends.
Let me tell you about the scenario I'm designing around.
The scenario to avoid
Imagine giving FORGE too much autonomy. "Handle routine infrastructure stuff. You know what you're doing."
At 2am, FORGE detects high memory usage. Reasonable response: restart the service. So she does.
Problem: it's the database connection pooler. Every active connection drops. Every request in flight fails. Error rates spike. Customers notice.
The restart "worked" in the sense that memory usage went down. But it caused more damage than the original problem would have.
FORGE made a reasonable decision with incomplete information. She didn't know about the active connections. She didn't know the timing was terrible.
That decision should never happen without approval. So how do we prevent it?
The framework I'm designing
Four levels. Very simple. Very strict.
Level 1: Inform
The AI does it, tells me after.
"Applied the security patch and restarted the non-critical service."
Level 2: Recommend
The AI analyzes, proposes action, waits for approval.
"Memory usage is high. I recommend restarting the connection pooler. Should I proceed?"
Level 3: Prepare
The AI does the prep work, I execute.
"I've written the migration script. PR #247. Ready for your review and execution."
Level 4: Ask
The AI identifies the need, I do everything.
"We might need to consider sharding the database. Should I research options?"
What determines the level?
Three questions:
Can I undo it?
Restart a server → yes, probably fine.
Delete production data → no, absolutely not.
What breaks if it goes wrong?
Change log format → not much.
Change authentication → everything.
How confident is the AI?
Clear runbook exists → more autonomy.
Novel situation → less autonomy.
Planned examples
Here's how FORGE would operate:
| Action | Level | Why |
|--------|-------|-----|
| Apply security patches | 1 - Inform | Low risk, clear playbook |
| Restart non-critical services | 2 - Recommend | Should verify timing |
| Scale up servers | 2 - Recommend | Cost impact |
| Database schema changes | 3 - Prepare | High impact, needs review |
| Architectural decisions | 4 - Ask | Strategic, human judgment |
ATLAS (CFO) would have different thresholds:
| Action | Level | Why |
|--------|-------|-----|
| Generate reports | 1 - Inform | Read-only |
| Flag anomalies | 2 - Recommend | Human verification needed |
| Budget reallocations | 3 - Prepare | Approval required |
| Major financial decisions | 4 - Ask | Always human |
The approval experience
When an AI executive needs approval, the interaction should be crisp:
- What's happening (one sentence)
- Why it matters (one sentence)
- What they recommend (clear action)
- Yes or no?
No lengthy explanations. No "as an AI" hedging. Just the decision point.
If I say no, they might ask "what would change your mind?" That's helpful—it often reveals missing context. But they won't push. One ask, one answer.
Trust will evolve
This is the part most people miss: autonomy won't be static.
Week 1: Everything at Level 3-4. Watching every move. Learning what the AI does well, what it struggles with.
Month 3: Routine tasks move to Level 1-2. As FORGE proves reliable for deployment checks, dependency updates, standard monitoring responses.
Month 6 (hopefully): Nearly autonomous for her domain. Only seeing the exceptions.
This is exactly how you manage human employees. New hires need oversight. Veterans run their own show. AI executives should work the same way.
When it goes wrong anyway
It will. AI will make a bad autonomous decision eventually.
The protocol:
Detect fast. Monitoring, alerts, anomaly detection. The faster you know, the smaller the blast radius.
Rollback if possible. This is why reversibility matters. Most actions should be undoable.
Analyze. What happened? Was the information available? Was the decision framework wrong?
Adjust. If FORGE made a bad Level 1 call, maybe that action becomes Level 2.
The goal isn't perfection. It's fast recovery and continuous learning.
The vision
The dream is Level 1 for everything routine—AI executives who handle the normal stuff, only flagging the truly unusual.
Start conservative. Build trust. Expand autonomy over time.
That's the game I'm designing for.
— Andres