AI Makes Great Developers Dangerous
AI did not simplify software. It gave experienced developers far more leverage and exposed how often companies confus...
Crew Resource Management for AI Pairing
7 min read
Talk Like a Cockpit, Not a Brainstorm
24.07.2026, By Stephan Schwab
Most AI pairing fails for the same reason bad cockpits fail: vague roles, sloppy confirmations, and too much polite guessing. Aviation fixed that decades ago with Crew Resource Management. Not with motivational posters. With standard phrases, explicit handoffs, challenge-response checklists, and the expectation that the second person speaks up before the metal bends. That discipline looks repetitive from the outside. Good. Repetition is cheaper than wreckage. Developers working with AI agents need the same posture, especially when the agent is fast, plausible, and wrong in ways that look tidy on first inspection. If you want better results from AI pairing, stop treating the conversation like improv. Treat it like a cockpit. Decide who is flying, who is monitoring, what phase you are in, what change is being made, how the other side must read it back, and what evidence counts as “complete.” You do not need aviation cosplay. You need less ambiguity. Precision feels slower right up to the moment it saves you a week of rework. The teams that figure this out early will look slower to people addicted to prompt theatrics. They will also ship with fewer stupid wounds. What would change if your team treated every AI handoff as a cockpit callout instead of a chat?
The Repetition Is the Safety Feature
In a serious cockpit, repetition is not inefficiency. It is error control.
Listen to two pilots who know what they are doing and the conversation sounds almost insultingly basic. One says the thing. The other repeats the thing. The first confirms the repeat. Checklist items are called, answered, and verified. Transfer of control is spoken out loud. Nobody is trying to sound creative.
That is not because pilots are dim. It is because ambiguity loves speed, stress, and ego. Cockpit discipline exists to make the right thing easy to hear and the wrong thing awkward to miss.
Most AI pairing does the opposite. The human hints. The model infers. The result arrives wrapped in confidence. Everyone pretends that “looks good” is a control mechanism. It is not. It is vibes wearing a lanyard.
As with When AI Becomes Your Thinking Partner, the value is in the conversation. The difference is that good conversation under load needs a protocol, not just intelligence.
Three Cockpit Habits Worth Stealing
Positive Transfer of Control
If nobody says who is flying, everybody assumes somebody competent must be flying. That is how messes get airborne.
The FAA teaches a three-step transfer of flight controls: one pilot hands off, the other accepts, the first confirms the transfer. That sounds pedantic until you imagine the alternative.
Use the same move with AI.
Do not say:
Can you clean this up a bit?
Say:
You are drafting. I am deciding.
Target: extract the billing validation logic from checkout.ts into a pure function.
Do not edit tests yet.
Read back the intended change before touching code.
The agent should answer with a readback, not with instant output:
Readback: I will propose a pure function extraction in checkout.ts, leave tests unchanged,
and wait for approval on the shape before broader edits.
That one exchange removes half the usual nonsense. Scope is explicit. Authority is explicit. The phase is explicit. Nobody has to guess whether the agent is exploring, proposing, or executing.
Challenge, Response, Verification
Checklist culture works because “done” is not a feeling. It is a verified state.
FAA guidance for flight deck crews describes normal checklists as a challenge-response process. One crewmember calls the item. The other verifies and answers. The point is not ceremony. The point is to stop memory and momentum from impersonating reality.
That maps directly to AI pairing:
Human: Challenge: add a feature flag around the new login path.
Agent: Response: feature flag added only at the HTTP entry point, default false.
Human: Verify: show me every file touched and explain why no downstream auth flow changed.
Notice what happened there. “Add a flag” was not the end of the exchange. The response included scope. Verification demanded evidence. That is the part teams skip when they are drunk on output.
If the change is risky, force the agent to answer four things every time:
- Intent
- Files touched
- Assumptions
- Verification performed
That is your lightweight cockpit checklist. Boring on purpose.
Readback and Hearback
Readback without hearback is theater. The sender has to notice when the reply is wrong.
NASA’s ASRS has been warning about readback/hearback failures for years: the receiver repeats the instruction incorrectly, and the sender fails to catch it. Aviation learned the ugly truth. Communication does not succeed when words are spoken. It succeeds when the other side understood the right thing and somebody verified that they did.
That matters with AI because the model often restates your request in prettier language while quietly changing the meaning.
Example:
Human: Migrate only the admin endpoints to the new auth middleware.
Agent: Readback: I will update the authentication middleware usage across the service.
That is not a harmless paraphrase. That is scope creep disguised as helpfulness.
The correct response from the human is not “close enough.” The correct response is:
Negative. Admin endpoints only. Name the exact routes you will touch.
Pilots do not get bonus points for being agreeable when the clearance is wrong. Neither should you.
A Simple SOP for AI Pairing
You do not need a giant process deck. You need five standard moves that show up every time.
If you want a usable operating procedure, start here:
- State the phase.
- Assign roles.
- Require a readback.
- Execute one bounded change.
- Confirm evidence before moving on.
In practice it can look like this:
Phase: investigation
You are monitoring the repository and proposing options. I am deciding.
Goal: understand why duplicate invoices are created.
Read back your current theory in three bullets before suggesting a fix.
Then later:
Phase: execution
You are drafting. I am approving.
Change: add idempotency guard in InvoiceService#create.
Constraints: no schema change, no background job changes.
Before completion, report touched files, tests run, remaining risks, and rollback plan.
This is not bureaucracy. This is how you keep children of the magenta line from blindly following whatever the automation drew on the screen.
Where This Pays Off Fast
The payoff shows up first in places where AI failure is expensive but not immediately obvious:
- Refactors with wide blast radius
- Authentication and authorization changes
- Data migrations
- Infrastructure edits
- Test rewrites that can accidentally bless broken behavior
These are exactly the situations where the model’s speed becomes dangerous. Fast output compresses the time available for noticing drift. Standard callouts buy that time back.
They also reduce a quieter problem: human fatigue. When the protocol carries some of the cognitive load, you spend less energy reconstructing what the agent thought it was doing. That matters more than most teams admit.
The Point of CRM
Crew Resource Management was never about making pilots sound formal. It was about making human coordination hold up when attention narrows, workload spikes, and mistakes become expensive.
That is not just an aviation problem anymore.
Pairing with AI agents gets better the moment you stop treating every prompt like a clever one-off and start treating the interaction as an operational handoff between two actors with different strengths and different failure modes. One is fast, tireless, and fluent. The other is accountable. If that accountability stays vague, the pairing stays sloppy.
Steal the boring parts from aviation. Explicit control transfer. Challenge-response. Readback and hearback. Completion signals. Standard phrases.
The cockpit already did the experimentation for you. You can keep improvising if you want. The incident report will also be custom.
Aviation Basis
- FAA AC 61-98E describes a three-step transfer of control with verbal handoff, acceptance, and confirmation.
- FAA AC 120-71B describes challenge-response checklists, clear initiation cues, and explicit completion calls.
- NASA ASRS documents the persistent readback/hearback failure pattern in pilot-controller communication.
Sources:
Talk It Through
Tell me what is happening. I listen, ask a few practical questions, and reflect back what I see: where the risk may sit, what may be blocking delivery, and what looks worth checking next. No pitch, no obligation. Confidential and direct.
Talk it through. Practical reflection, no pitch.
Start a Conversation
🎭 This Week's Episodes
Telenovelas show what we can't say in client meetings. The drama is heightened, but the patterns are real.
The New Baseline
With the tournament system successfully launched and the network partition survived, the team gathers for a Q3 retrospective. Katja and Stefan pres...
The First Confrontation
Nathan Cole thinks the Detroit tax defect has finally become visible enough to be handled like an adult problem: evidence on the screen, a bounded ...
Technical Consultancy
Embedded Delivery Partner
Embedded into your team as an active contributor, reducing delivery friction and helping important work move cleanly.
Technical Advisor
Peer-style technical assessments before far-reaching decisions; reduce architectural and product risk early.
Product & Delivery
Ship working software to real users earlier. Measure impact and adapt based on evidence rather than assumption.
Custom Software Development
High quality, maintainable software. Short-term augmentation that leaves lasting capability in your own team.
Recent Articles
- Crew Resource Management for AI Pairing (24.07.2026)
- Agentic Coding Is Not 'Make Me X' (20.07.2026)
- AI Agents Make You a Software Company (16.07.2026)
- You Never Needed a Fad to Invest in Your People (11.07.2026)
- Agentic Coding Needs an Air Traffic Control Shift Model (07.07.2026)