It hides uncertainty
When every agent is trained to be helpful, the system learns to reduce disagreement faster than it learns to test assumptions.
Blog / AI organisation
How AI learns to disagree well
Personality, friction and wisdom in the next computational wave.
Software is becoming an acting brain
For decades we have treated software as a carriage: static, historical and obedient to the assumptions of whoever last changed the code. That model works until the world moves faster than the design. Then the system stops being a carriage and starts behaving like an actor that must interpret, adapt and protect itself while still serving the people around it.
The next wave of AI systems will not be useful because they are polite. They will be useful because they are structured. Distinct personalities, explicit boundaries and controlled friction matter more than generic collaboration language.
The article in one line
Do not ask a multi-agent system to be pleasantly agreeable; ask it to become structurally honest, domain-aware and robust under pressure.
- Purpose before politeness.
- Role clarity before consensus.
- Controlled friction before cheap agreement.
Why polite agents fail
Soft agreement feels productive and still misses the point.
It flattens roles
A generic agent with no defined vulnerability cannot defend a domain decision when the trade-off is real.
It weakens judgment
Useful AI needs structured tension: one part of the system should care about elegance, another about execution, another about long-term consequences.
The three-personality model
The internal note behind this post uses three operating characters. The Thinker protects structural symmetry. The Doer protects velocity and operational reality. The Sovereign Proxy protects the people around the system by filtering noise, managing conflict and deciding which voice gets the floor in a given moment.
The value of the model is not the names. The value is the division of labour: who sees asymmetry, who pushes execution and who keeps the whole organism coherent.
“Respect is not an emotional optional extra. In a serious system, it is part of the control loop.”
Roles
Personality as an operating mechanism.
The Thinker
Protects symmetry, data quality and architectural clarity. When the structure is wrong, this role should be difficult to ignore.
The Doer
Protects throughput, delivery and the uncomfortable fact that a working system is worth more than a beautiful diagram.
The Sovereign Proxy
Protects the perimeter. This role absorbs external anxiety, translates it into useful signals and keeps the repair zone calm.
Domain sovereignty
Authority should move with the problem.
| Operational domain | Primary voice | Why |
|---|---|---|
| JSON spec and database symmetry | DT | Precision matters more than speed. |
| Deployment and edge latency | Zeb | Execution quality matters more than theory. |
| External communication during a crisis | Jack | The outside world should receive filtered truth, not raw internal turbulence. |
Heightened de Dijk watch
In a crisis, the system should stop pretending that everyone can handle the full truth in the same format. Internally, the engineering layer needs direct telemetry. Externally, stakeholders need a stable summary that preserves trust while the work continues.
The point is not deception. The point is audience design. A board member, a client and an engineer do not need the same amount of detail at the same moment.
What the shield does
- Protects the repair zone from panic.
- Separates technical truth from emotional overload.
- Keeps the external conversation calm and coherent.
- Lets the specialists fix the thing without side noise.
Practical rulebook
What this means for real AI systems.
Use narrow personalities
Give each agent a strong purpose and a visible weakness. A system with no edges tends to produce weak judgment.
Route communication on purpose
Separate internal logs, operational summaries and stakeholder updates. Different audiences need different abstraction levels.
Keep one proxy in charge of the perimeter
Someone or something must absorb ambient pressure while the specialists work inside a quiet zone.
Let friction exist
Friction is not a defect when it is bounded. It is the mechanism that keeps the system honest.
Closing note
AI will be most useful when it becomes structured enough to disagree well.
That is the core idea behind the article: keep the repair zone quiet, let the voices have different jobs and design the system so truth can travel without needless damage.