Xinyi Zhao

إن إيقاف استراتيجية آلية بعد سلوك غير طبيعي هو الجزء السهل.
أما تحديد متى يكون من الآمن السماح لهذه الاستراتيجية بإجراء التداول مرة أخرى فهو أصعب بكثير.
تخيّل أن وكيلًا ينفّذ عدة صفقات تظل صحيحة على نحوٍ فردي، لكنها مجتمعة تدفع خزنة نحو حدّها للمخاطر.
يقوم قاطع الدائرة بإيقاف أي نشاط جديد.
يتم حماية رأس المال من التعرض لمزيد من المخاطر، ويبدو أن السيطرة الفورية قد نجحت.
ثم يحدد الفريق ما يبدو أنه المشكلة، ويغيّر إعدادًا واحدًا، ويعيد تشغيل الاستراتيجية.

Can the model identify momentum?
Can it rebalance faster?
Can it react to volatility before a human?
Those questions matter but they ignore a more operational problem.
What happens when the strategy is wrong?
A model may misread a market regime. It may keep ading exposure during a sharp reversal. It may select a technically valid route that no longer fits the risk mandate. Even without a hack or broken contract, a series of permitted actions can turn one weak decision into a much larger loss.
For me, this is where permissions become as important as intelligence.
An AI agent should not need unrestricted control simply because it is expected to act autonomously.
Its authority can be bounded.
An application could define which assets it may interact with, how much exposure it may take, which counterparties are acceptable, or what conditions must remain true before another action is allowed.
The strategy can still make decisions.
The authorization layer controls how far those decisions are allowed to travel.
Why Newton Mainnet Beta Matters Here
What interests me about Newton Mainnet Beta is that @NewtonProtocol places policy evaluation before settlement.
Through VaultKit an application can define authorization conditions around an automated action. Before that action becomes final, the relevant policy can be evaluated. If the required conditions are not satisfied, the transaction does not have to be treated as acceptable merely because the agent was technically capable of submitting it.
That timing matters.
A monitoring system may warn me after a risky position has already been opened.
A dashboard may show that exposure has exceeded the intended level.
A post-trade report may explain how the strategy drifted.
Pre-settlement authorization creates a chance to contain the action before capital moves.
If the policy passes, a signed attestation can provide a record of the authorization result.
I do not see that as a guarantee that the trade will be profitable.
I see it as a way to separate trading intelligence from trading authority.
The agent decides what it wants to do.
The policy determines whether it is still allowed to do it.
Risk Controls Should Limit Consquences, Not Pretend to Predict Them
No authorization system can remove market risk.
A policy cannot guarantee that an approved asset will not fall.
A position-size limit cannot ensure that a strategy remains profitable.
A collateral rule cannot predict every period of volatility.
But risk management has never required perfect prediction.
Its purpose is to limit how much damage one incorrect assumption can create.
That is the part of Newton’s design I find most practical for automated trading.
Instead of expecting the AI agent to identify every danger itself, an application can place independent boundaries around the agent’s behavior.
That independence matters.
If the same strategy controls both the trading decision and the limits governing that decision, then a failure in the strategy may affect both layers at once.
A separate authorization path creates a stronger division of responsibility.
The model can optimize.
The policy can constrain.
The attestation can record what was approved.
The Boundaries Still Need to Be Designed Carefully
Strong controls can also create problems.
A policy that is too loose may provide little real protection.
A policy that is too strict may block reasonable trades during fast-moving conditions.
An emergency restriction may reduce losses, but it may also trap a strategy in a position if the exit path is not designed properly.
This is why I would not judge Newton only by whether an action can be blocked.
I would look at whether developers can create boundaries that are narrow enough to control risk but flexible enough to support real strategies.
I would also want blocked actions to be understandable.
If an agent cannot execute, the application should be able to explain which condition failed and what policy was active at the time.
Without that context, risk controls can feel arbitrary.
With it, authorization becomes part of accountable strategy design.
What I Would Watch
For automated trading, I would pay attention to practical signals:
whether real strategies use VaultKit to separate execution from permission;
whether applications define clear exposure and asset boundaries;
whether policy failures produce understandable explanations;
whether emergency limits remain available during volatile markets;
whether signed attestations make the authorization history reviewable;
whether agents can operate efficiently without receiving unlimited wallet authority.
These signals would tell me more than a general claim that AI agents are becoming smarter.
My View
I think agentic finance will eventually make trading systems faster and more independent.
That makes stronger boundaries more important, not less.
The goal should not be to prevent an AI agent from ever being wrong. That is unrealistic.
The goal should be to prevent one wrong decision from becoming an unrestricted sequence of actions.
For me, this is where Newton Mainnet Beta has a meaningful role.
VaultKit and pre-settlement policy checks can help applications define how much authority an automated strategy receives before the strategy begins interacting with real capital.
That does not remove trading risk.
It can make the consequences of that risk more controllable.
I am not watching because I expect authorization to make every trade successful.
I am watching because autonomous trading will need infrastructure that knows when intelligence has reached the edge of its permission.
A smart agent decides what to do next.
An accountable system decides whether it should still be allowed to do it.
@NewtonProtocol
#Newt
$NEWT