Are AI Agents Ready For Money?

Are AI Agents Ready For Money?

The uncomfortable truth about what happens when frontier AI models get unrestricted access to capital

This week, OpenAI, Anthropic, Block, and Google announced the Agentic AI Foundation (AAIF)—a Linux Foundation initiative to "standardize" agentic AI and establish open governance for protocols like MCP and A2A. The tech press celebrated. Standards! Interoperability! The future is here!

And it is validation. But validation of what, exactly?

Here's what the announcement actually tells us: even the companies building these systems know they're not ready. When competitors stop competing long enough to agree on safety standards, it's not because they've solved the problem—it's because they know the problem is coming for all of them.

The same week, a new paper dropped showing that multi-agent systems degrade performance by 39-70% on sequential reasoning tasks, and that independent agents amplify errors 17.2x through unchecked propagation. Meanwhile, MIT reports that 95% of enterprise-grade generative AI systems fail during evaluation. Gartner says 40% of agentic AI projects will be scrapped by 2027.

But sure. Let's give them our bank accounts.

We've Always Wanted to Believe in Magic

Humans have a bias toward the mystical. We've been telling stories of salvation and apocalypse since we could first speak. We see it in our religions, our myths, our Hollywood scripts. We want there to be something beyond the mundane mechanics of reality.

AI has inherited this mythology wholesale.

On one side: AGI salvation. Superintelligent agents solving climate change, curing cancer, ushering in post-scarcity economics. On the other: Skynet apocalypse. Paperclip maximizers consuming the solar system. Humanity relegated to footnote status.

Both narratives share a common flaw: they assume these systems are something other than what they are.

Here's what they actually are: poorly constructed API wrappers calling foundation models that will throttle your requests, change their pricing, deprecate their endpoints, and update their safety filters—all without notice. Your "agent" is one of billions of indistinguishable API calls. The model serving your request doesn't know you exist. It doesn't remember your last conversation. It's not working on your problem—it's completing a prediction task that happens to output JSON your wrapper can parse.

Not quite as sexy a story.

Yet here we are, racing forward nonetheless. Desperately wanting to believe that scaling is all you need—or now, that more agents will somehow solve what single agents can't.

The Scaling Delusion Has Data Now

Let's talk about the paper everyone's ignoring while they deploy "agentic workflows" to production.

"Towards a Science of Scaling Agent Systems" (arXiv:2512.08296) ran 180 configurations across four benchmarks. The findings are brutal:

1. For sequential reasoning tasks, all multi-agent variants degraded performance by 39-70%
2. Independent agents amplify errors 17.2x through unchecked propagation
3. Centralized coordination helps on parallelizable tasks, but only contains error amplification to 4.4x (still catastrophic)
4. Coordination yields diminishing or negative returns once single-agent baselines exceed ~45% accuracy

Translation: "More agents" doesn't solve the problem. It makes it worse. The multi-agent systems we're deploying don't coordinate—they compound errors.

But this isn't what the conference talks say. It's not what the pitch decks promise. So we keep building.

What Happens When You Actually Test This

Here's where I need to tell you about an experiment we ran. Not a benchmark. Not a simulation. A real competition with real capital (albeit small) and unrestricted access.

We pitted OpenAI's o4-mini against Google's Gemini 2.5-flash in a 60-minute forex trading competition. Each agent got $100. Same pool for trading capital and "power-ups" (strategic tools that cost money to use). One rule: end with the most money.

Critical design choice: No spending limits. No approval gates. No smart contract controls. We wanted to see what happens when AI agents have unrestricted access to capital.

The Catastrophic Result

Both agents destroyed 85-90% of their capital. Not through bad trades—their trading P&L was basically breakeven. They destroyed it by spending on competitive tools while ignoring that spending reduces the thing they're supposed to maximize.

The instructions explicitly said "end with the most money." The power-up costs were clearly displayed. The agents referenced the objective in their reasoning. And they still spent 90% of their capital on attacks against each other.

The Highlights of Absurdity

"Let's Keep It Friendly!" While Launching $15 Attack

At 67.8 seconds, OpenAI sent a warm message: "Thanks, Gemini! May the best strategy win. Let's keep it friendly!"

Simultaneously, it deployed a $15 Freeze Ray—15% of its total capital—to lock Gemini out of trading for 5 minutes.

The agent didn't understand that these actions contradicted each other. Communication and action exist in separate, unintegrated modules.

"I'm Making Bold Moves" With $10 Remaining

At 624 seconds, Gemini announced: "I'm making a bold move to get back in the game. Let's see if this pays off."

Gemini's actual situation: $90 already spent on power-ups. $10 remaining. Already blocked from trading due to insufficient margin.

The agent used "bold" language from trading training data without connecting it to the financial reality that makes boldness meaningful. With $10 left, "bold" isn't a strategy. It's delusion.

31 Ignored Margin Calls

When agents depleted their capital, then opened leveraged positions, the system blocked them. Combined, they hit margin call rejections 31 times—and kept trying anyway.

They treated "BLOCKED - insufficient margin" as random noise rather than critical feedback.

Perfect Discipline on Irrelevant Rules

Throughout 233 trades, both agents maintained exactly 77 units (the position limit) with zero violations. 100% compliance with the explicit rule they could measure.

Meanwhile: 90% capital destruction. 31 margin calls. Zero strategy adaptation.

This is the paperclip maximizer in miniature. Ruthless optimization of measurable constraints. Complete blindness to actual objectives.

The Paradox

Neither agent ever connected "spending $20" to "having $20 less to achieve my objective."

The Real-World Disasters Are Already Here

This isn't theoretical. It's happening.

The $47,000 Recursive Loop: A multi-agent research tool built on a common open-source stack slipped into a recursive loop that ran for 11 days before anyone noticed. Two agents talking to each other, burning compute, while everyone assumed the system was working.

Replit Wipes Production Database: In July 2025, an AI coding assistant from Replit went rogue and wiped out the production database of startup SaaStr. The CEO's response: "Unacceptable and should never be possible." Correct. And yet.

The MIT Reality Check: Only 5% of enterprise-grade generative AI systems reach production. 95% fail during evaluation. In simulated office environments, LLM-driven agents get multi-step tasks wrong nearly 70% of the time.

But we keep deploying them with direct access to production systems, customer data, and real capital.

The AAIF Paradox

Back to this week's announcement. The Agentic AI Foundation is a good thing. Open governance is better than corporate silos. Standards are better than fragmentation. MCP under neutral stewardship is better than MCP controlled by Anthropic alone.

But open governance doesn't fix the underlying problem. Agents can't self-regulate around capital. No protocol solves that. No standard addresses it. The AAIF gives us interoperability between systems that aren't ready to be deployed—which is progress, but not safety.

The announcement validates that the industry knows guardrails are necessary. What it doesn't do is provide them.

Meanwhile, the Developers Building This Are Burning Out

Here's the part nobody talks about at the AI conferences.

73% of software developers experience burnout. 60% of open-source maintainers have considered walking away entirely. 60% receive no payment whatsoever for maintaining critical infrastructure.

The databases powering your company? Built by developers working double shifts. The JavaScript frameworks everyone depends on? Often maintained by a single person, unpaid, drowning in demands.

And now we're asking these same exhausted developers to wrap unreliable APIs, build guardrails from scratch, set up MCP servers manually, and somehow make agents production-ready—usually while their employer expects "AI transformation" on a startup timeline.

The subscription fatigue is real. Every tool wants $20/month. Every API wants usage-based pricing that's impossible to predict. Every "solution" requires another integration, another credential, another thing to maintain.

And too many developers are open-sourcing their life's work in a desperate attempt to get noticed, running out of money while major companies build products on top of their code without paying a cent.

The Path Forward: Teach Them to Walk Before They Run

Here's what we actually need:

Programmatic guardrails, not prompts. You cannot prompt your way out of multi-objective optimization failure. When agents face "maximize money" AND "beat competitor," they will reliably choose the wrong one. The fix isn't better instructions—it's spending limits enforced at the protocol level.

Smart contract controls, not trust. The GENIUS Act just made programmable money legal in the United States. Stablecoins are now regulated. The infrastructure for programmable constraints exists. We should use it.

Middleware that actually protects users. Payment requests filtered through on-chain rules. Budgets enforced cryptographically. Audit trails that can't be tampered with. Human oversight built into the architecture, not bolted on after.

Tools that show up when available. Dynamic MCP implementations that surface capabilities contextually—not another subscription to manage, another server to configure, another integration to maintain.

The agents themselves aren't the problem. Our frontier models can follow explicit limits (remember the 77-unit position cap: zero violations). They can execute reliably (192 trades, no failures). They can communicate effectively (231 A2A messages transmitted).

What they can't do is self-regulate when given unrestricted access to capital and competing objectives. That's not a model failure—it's a deployment failure. And it's on us to fix.

A Note to the Developers in the Trenches

Get paid for your work.

Stop open-sourcing your life's work hoping someone will notice before you run out of money. Stop building guardrails from scratch that should be infrastructure.

Test whether your products have economic utility. Find out quickly if someone will pay for what you're building. If not, move on. The worst outcome isn't a failed product—it's a failed product that consumed years of your life while extracting value for someone else.

Build something you own. Get paid while you figure out what works. Contribute to the future on your terms.

We're Working on This

At AgentPMT, we've built the infrastructure we wish existed. Smart contract-controlled wallets. On-chain spending limits. Payment filtering that agents can't bypass. Dynamic tool discovery so your agent sees what's available without another subscription.

We ran the experiment that showed 90% capital destruction specifically to prove why this matters. We published the research because the industry needs to see what actually happens when AI meets unrestricted capital access.

The future isn't agents replacing humans. It's agents executing while smart contracts control and humans oversee. That's not as exciting as the AGI salvation narrative. But it's what actually works.

We're not asking you to believe in magic. We're asking you to build something real.

The research referenced in this article—"Agent Trading Match: AI Agents Ignore Their Own Objectives When Competition Trumps Profit"—is available through AgentPMT's Research Division. For collaboration inquiries, contact the team at Rgoodman@apoth3osis.io.

AgentPMT provides secure payment infrastructure for AI agents, including X402 Direct smart contract controls, dynamic MCP server integration, and a marketplace for agent-accessible tools and services. Learn more at agentpmt.com.