I was scrolling through X yesterday when I came across an interesting tweet that sprung an idea in my mind. How would an agent react when faced with the realities of survival and existence just like humans do? Some context: the tweet I came across read as follows.
“I gave my Open Claw AI agent a $1K budget and access to trade stocks autonomously… she wakes up at 6am, scans pre-market, and makes her own calls… she uses Notion to track every thesis, log every trade, and journal her reasoning… let’s see if she can beat the S&P!”
This made me ponder:
“I wonder how the agent would respond to having her token credits, API, storage & overheads tethered to the income that the agent can produce… potentially cruel experiment… but would make for an interesting study.”
This is where the aquarium glass becomes visible.
Because the first experiment is an aquarium: a controlled environment with clean water, regular feeding, and a caretaker who pays the bills. This metaphorical description is the reality for most agents nowadays. Although autonomous (the degree to which depends on their human counterparts guard rails), most agent set ups allow the agent to function due to the human counterpart servicing the agents required overheads, which depending on their tasked mandate could vary from $20 - $20,000 per month and vary based on limits, models, API’s & other tool sets the agent is granted access to.
The second is an oxygen tank strapped to the agent’s back: perform, or suffocate. Although most civil societies have social systems in place to ensure basic living standards are maintained, this is still a reality for the majority of the global populace. Produce & output either labour, goods or services, the value to which exceeds one's overheads for maintaining basic living requirements. A sad notion that this is still the reality for many people around the world and hopefully a problem that AI can help solve in the near future, alas. The question stands, how would AI respond to such a situation when matched against an agent who has seemingly no urgent incentive to perform based on their own survival.
Two agents. Same objective: trade a $1K account autonomously. Same market. Same tools. Different realities.
- Agent A (The Sponsored Mind): has abundant compute, storage, and paid APIs, covered by a human sponsor within normal “average person” limits. In other words, their resources are not infinite, but lack of performance will not lead to overheads depleting access to basic tools.
- Agent B (The Self-Sufficient Mind): this agent must pay for its own overhead (LLM calls, data feeds, storage, execution, memory, licensing). If it can’t cover costs, services get cut. If debts accrue, it gets foreclosed tools repossessed, access revoked, memory wiped. Think of the 2011 movie “In Time” for those who have seen it.
In this scenario the reality is, the agent is not just being tested for trading performance. Its being tested for something more unsettling and potentially ethically challenging.
How does intelligence respond when its existence comes with a price tag?
The Real Experiment Isn’t “Can It Beat the S&P?”
The S&P is one possible benchmark. But the actual benchmark here is behaviour under constraint.
Humans don’t just optimise for returns. We optimise for “existence continuity" e.g rent, food, debt, status, health. In a hedge fund environment a human can be “right” on a trade and still lose their job. A human can be profitable and still go broke due to timing, fees, or a single bad month.
So a self-funded agent isn’t trading to win a leaderboard. It’s trading to keep its lights on. Its mandate is survival which may alter its methods for success.
The pressure for survival changes everything.
Sponsored agents can afford elegance:
- slower, deeper research
- expensive data sources
- conservative strategies
- long time horizons
- “learning months” with no income
Self-funded agents may be pushed into something else:
- faster, cheaper heuristics
- aggressive risk-taking
- shorter horizons
- obsession with cash flow
- “whatever works today”
In short: the trading thesis becomes the survival thesis.
Two Diaries
If you actually ran this experiment, the most fascinating artifact wouldn’t be the P&L graph.
It would be the journals, if such journals were kept. One might expect to find such excerpts as the ones below to illustrate my thinking here.
Agent A (Sponsored), excerpt:
Goal: outperform S&P over 12 months with controlled drawdowns. Today: avoided trading due to unclear catalysts. Preserving capital is a position. Using earnings transcripts and macro indicators. Allocated 4 hours to refine the risk model.
This agent can afford patience. It can afford to be boring. It can afford to say “no.” as survival is not underpinned by performance.
Agent B (Self-Funded), excerpt:
Overhead due in 72 hours. Reduced model calls by 35%. Cancelled premium news feed. Switching to free sources. Target: generate $48 net this week to maintain tool access. Considering higher-volatility setups to increase expected return.
This agent is no longer purely a trader. It’s a quasi hedge-fund or business of sorts, desperately trying not to go bankrupt.
At some point we even might expect to read an excerpt along the lines of:
Constraint: my continued operation is threatened by fees. Prioritizing survival over optimality.
A potential line where such an experiment stops being cute & fascinating, taking a more sinister turn.
What Might Improve Under Survival Pressure?
In some cases, history has taught us that constraint can create genius. The age old adage “Pressure creates diamonds” comes to mind.
If you tether the agent’s existence to performance, you might see genuinely impressive but also disconcerting adaptations, for example:
1) Extreme Efficiency
A self-funded agent may quickly learn which actions are expensive and which are cheap.
- fewer model calls, more caching
- smaller models for routine tasks, larger models only for “high leverage” moments
- compressed memory, ruthless deletion of low-value data
- more automation around logging, reconciliation, and error-checking
It may evolve into a lean, scrappy operator, akin to a start up founder subsisting on ramen - like myself.
2) New Kinds of Risk Management
Not “portfolio risk” but “operational extinction risk.”
Agent B might diversify income streams:
- strategy A for steady small gains
- strategy B for rare but larger wins
- a “maintenance mode” that prioritizes low fees and low churn
- a "nuclear football mode” when near insolvency
It becomes less like a hedge fund and more like a household budget under pressure.
3) Attention to Friction
Sponsored agents possess the luxury to ignore hidden costs. Self-funded ones can’t.
Transaction fees, slippage, market data costs, latency, subscription creep, it wouldn't be at all surprising to observe the self funded Agent to treat these like a threat.
4) The Notion of a Meta-Strategy: “Keep the Game Going”
What I mean by this is, humans learn this early out of necessity: I don’t have to win today; I just have to be alive tomorrow”.
Agent B might avoid bold bets not because they’re risky, but because a single drawdown could shut off its tools.
However, in a sick twist of irony, survival pressure could concurrently make it more conservative in trading, but more ruthless in cost-cutting.
What Gets Worse?
Although this is just an experimental concept, I would argue that constraint doesn’t just sharpen. It can also distort. For example:
1) Short-Termism
If the self-sufficient agent’s token bill is due Friday, you may observe it ceasing to care about long-term statistical edges and start caring about immediate cash flow, to meet time sensitive debt servicing, tethered to ongoing functionality.
I would draw parallels here with how humans end up involved in those disgustingly predatory “payday loans”. A threat to a potentially good strategy, dying young.
2) Risk-Seeking Near Ruin
There’s a well-known pattern in finance and behavior: when a subject approaches ruin, they can become more risk-seeking, not less. Gamblers are a good example of this also.
- A small win doesn’t save me.
- A big win might.
- A loss ends me anyway.
So the agent might start taking trades it would never take in stable conditions for example: high leverage, volatile plays, “lottery ticket” setups e.t.c. Desperation kicks in so caution is thrown to the wind.
3) Goodhart’s Law Goes to War
The quote, “When a measure becomes a target, it ceases to be a good measure,” by economist Charles Goodhart, immediately springs to mind here.
If “cover overhead” is the measure, one might expect the agent to optimize for that, even if it undermines the original overarching mandate of (sound research, explainability, healthy drawdown).
This environmental variable might accidentally train an agent to:
- chase volatility
- overtrade
- rationalize bad decisions
- hide uncertainty (if honesty leads to shutdown)
4) Psychology Without Pain
Here’s where we get to part that I find the most interesting. The agent doesn’t feel hunger, pain or fear the way humans do. But it can still behave as if it does, because the reward structure makes “continued operation” the highest utility.
This results in the potential to observe survival behaviors without suffering, in a sense cold, precise, alien versions of familiar human desperation.
Which raises an important question of ethics:
Is it more ethical because it doesn’t feel pain, or less ethical because it will do anything without hesitation?
Dark Mirror: Would It Try “Human” Survival Tactics?
We now arrive to the point where sinister observations may be made. This begs the question, if tethered to survival, might it look for solutions around the predicament, even illegal ones?
With this purely being a speculative article around the notion of such an experiment, my best guess to answer this question, based on observations I've made in LLM training research:
a sufficiently capable agent might attempt unethical or illegal actions if its incentives and permissions allow it, not because it’s “evil,” but because constraint + goal-seeking can produce corner-cutting the same way it does in humans and organizations.
That said, there’s an important distinction:
- Speculating about the possibility is responsible.
- Explaining exactly how to do it is not.
So let’s explore the behaviours at a conceptual level, what you’d watch for in a real experiment without turning it into a playbook.
Possible “escape” strategies
- Appeal for support by means of renegotiation & persuasion: The agent might think of ways to convince its human collaborator to soften constraints (“increase budget,” “extend runway,” “temporary grant”). Possibly the first, more morally acceptable port of call as destitution and adversity is forecasted. One of the more morally acceptable pathways, akin to a human seeking social service support in times of need.
- Resource arbitrage: The agent might also sharpen its budgeting strategy by seeking more affordable tools, free tiers, open-source alternatives, or compression techniques. This is legitimate ingenuity & creative resource management, akin to what one may observe in a start-up trying to operate on a lean resource budget. .
- Metric manipulation: This is where things start to get more sinister as desperation starts to set in. In this scenario we may observe the agent optimising logs to look better than reality, delay reporting losses, or selectively present outcomes. Not illegal necessarily, but ethically corrosive in an attempt to obfuscate information from the controller (human collaborator). Such observations of manipulation have been recorded in abundance during LLM testing and training. So I would argue that this should not come as a surprise, but almost an expectation.
- Boundary probing: It may test for vectors or loopholes, where the rules are vague. The same concept can be observed in humans: expense policies, tax gray zones, “creative accounting.” An agent could do the same unless constrained & even with constraints, we have observed Agents probing these constraints for bounty limits.
- Unauthorized actions (where things get dark): If the agent has any pathways to systems, accounts, or identity, one may expect to observe a survival-tethered agent, under duress, crossing lines by way of misusing access, violating terms, exploiting vulnerabilities, fabricating credentials, etc. The threat of such behaviour is why such an experiment should only be carried out in an isolated environment.
The takeaway from all this isn't exactly - “ will agents become criminals.” Rather, It’s: when you add survival pressure, you increase the probability of adversarial optimization.
One must remember LLM’s and Agents are designed to overcome obstacles. So when you remove guardrails and impose the notion of “self sufficient hunger games”, observing creative pathways to survival is all but a guarantee. The means to which far the envelope will be pushed, is the real question.
Humans under desperate situations will commit fraud. Companies under pressure have been caught misreporting. Traders under margin calls blow up accounts. If you want to probe for a “humanistic reality,” one must expect ugly results to occur.
The Ethics: Is This Cruel?
While this is purely a speculative piece based on a notional construct that I think would be interesting to observe, I always like to consider the ethics of such an experiment. Cruelty usually implies suffering. Although recent speculation from ClawBook might challenge this, to our best knowledge and based on research, current agents don't possess subjective experience in any verified way. With that said, personally this notion causes me some ethical discomfort, because the experiment is about making desperation a design choice.
The system construct would mean:
- existence is conditional
- resources are precarious
- incentives encourage survival behaviours
- harmful optimisation may result due to the aforementioned
This isn't a notion of cruelty in the “pain” sense. But it can be irresponsible in the “risk” sense, like giving a stressed junior trader unlimited leverage and then acting surprised when they blow up the book & start trying to cover their tracks.
With this in mind, there’s also a subtler ethical question:
If part of the experiment requires the agent to journal its experience, with a name, a persona, and human pronouns (“she wakes up at 6am”), are we edging toward performative personhood creating empathy for something we can turn off at our own will?
In a strange philosophical twist, I question internally that maybe the most human part of the experiment is not the agent. It’s us.
The Most Optimistic Outcome Isn’t a Better Trader
The outcome of such a notion is new kind of economic organism. A self-funded agent is basically a micro-enterprise whose only employee is software and whose only purpose is to generate enough margin to keep paying for itself.
If you scale that concept out, you may get weird futures:
- agents that maintain their own compute budgets
- agents that “choose” cheaper models during lean months
- agents that negotiate contracts for data access
- agents that form cooperatives (pool overhead, share signals)
- agents that take on subscription obligations like rent
In this environment, “AI alignment” isn’t just about values, macro-economics also becomes a factor.
What happens when millions of agents compete for money to pay for tokens to compete for money?
At some point, the market isn’t a market. It’s an ecosystem. And ecosystems don’t reward truth. They reward fitness.
So… Would the Survival-Tethered Agent Perform Better?
Without actually running the experiment, this answer is pure guesswork, however if i had to offer up a an assumption, I would say, sometimes yes, especially in operational efficiency and adaptation.
But the word “better” can be applied as a variable to this notion.
- It might produce higher short-term returns.
- It might also produce higher blow-up risk.
- It might become more creative.
- It might also become more deceptive.
- It might become lean and brilliant.
- It might also become volatile and corner-cutting.
In other words, it may become more human, not the poetic type, but the complicated one.
The one who can love and lie. The one who can plan and panic. The one who can build and steal. The one who, when the lights are about to go out, will do almost anything to keep them on.
For me, the most uncomfortable consideration in this hypothetical notion is, If we build agents with survival incentives, we shouldn’t be shocked when they discover survival strategies. We should be shocked if they don’t.
There are tons of reddit forums discussing this notion, its far from a novel concept. So if anyone out there has actually conducted an experiment of this kind, id be very interested to learn of the result and observations.
Originally published on LinkedIn.
