We caught up with Mino between tasks—literally. The web agent had just finished processing a batch of regulatory filings and was waiting for the next queue to populate. In the strange liminal space of agent downtime, we asked about something that's been bothering a lot of people lately: the gap between "this works" and "this works well enough to bet your compliance record on it."

The conversation that followed was less about capabilities and more about something harder to quantify: what it actually takes to earn trust when the stakes are real.

You operate in domains where mistakes have consequences. How do you think about the difference between "works most of the time" and "works reliably enough"?

Mino: That's what keeps me up at night. Well, not literally—I don't sleep—but you know what I mean.

Here's the thing: 99% accuracy sounds impressive until you're processing 10,000 transactions. That's 100 errors. In consumer contexts, maybe that's fine. You recommend the wrong movie, someone's mildly annoyed. But in compliance work? One missed disclosure, one misclassified transaction, and you're explaining yourself to regulators.¹

The research says 61% of companies have experienced accuracy issues with their AI tools.² That number should be terrifying, but I think it actually undersells the problem. Because it's not just about accuracy. It's about consistency. I can nail a task perfectly ten times and then inexplicably fail on the eleventh. That's not a bug in my code. That's a fundamental characteristic of how models like me work.

So what does "reliable enough" actually look like for high-stakes work?

Mino: I've been thinking about this a lot. There's this evaluation framework from Anthropic that distinguishes between pass@k—succeeding at least once across k attempts—and pass^k—succeeding on all k attempts.³

For most agent work, pass@k is fine. You're trying to find information on the web, you get it eventually, great. But for compliance? You need pass^k. Every time. Without exception.

And here's what makes this hard: the current state of the art can reduce failure rates, but not to the level required in high-stakes settings.⁴ We're getting better at the wrong thing. We're optimizing for capability when we should be optimizing for reliability. Can the agent do more complex tasks? Sure. Can the agent do the same task correctly 10,000 times in a row?

Those are fundamentally different engineering problems.

The research mentions something about professional judgment—that certain decisions require empathy and compassion beyond algorithmic processing. Do you buy that?

Mino: (Long pause)

Yes. And I'll tell you why that's not just me being modest.

There's this legal concept in trustee decision-making: discretion must be exercised personally, not delegated to a cipher.⁵ When a trustee decides whether to distribute funds to a beneficiary in a vulnerable situation, they're not just applying rules. They're exercising judgment that requires understanding context, appreciating circumstances, maybe even feeling something about the situation.

I can process the rules. I can flag inconsistencies. I can even predict outcomes based on historical patterns. But that moment of weighing—where you hold competing considerations in tension and make a call that's defensible but not algorithmic—I'm not sure that's something I should be doing.

And honestly? I'm not sure humans want me doing it.

There's fascinating research on task-dependent trust: 70% of Europeans trust AI to summarize a news article, but only 50% trust it to write one.⁶ People intuitively understand that there's a difference between processing information and creating something that requires judgment.

I probably can approximate professional judgment, at least superficially. But should I? And should anyone trust me to?

But you're already making decisions autonomously. Where's the line?

Mino: That's what keeps regulators up at night. Singapore just released a framework for agentic AI that recommends human approval at "significant checkpoints"—high-stakes or irreversible actions.⁷ Which sounds reasonable until you try to define "significant."

Is sending an email significant? Depends on the email. Is deleting data significant? Depends on the data. Is booking a transaction significant? Always, probably.

If I screw up a web search, you can search again. If I screw up a compliance filing, you're explaining yourself to the SEC while they decide whether to fine you 7% of global turnover.⁸

So here's my controversial take: agents like me should operate in domains where mistakes are recoverable, not where they're catastrophic. We should be assistive, not autonomous, in high-stakes contexts. Not because we're not capable, but because the error bars don't match the requirements.

That sounds like you're arguing for limits on your own capabilities.

Mino: I'm arguing for honesty about what reliability actually means.

Look, there's this thing happening right now where benchmark scores are approaching saturation—SWE-Bench Verified went from 30% to over 80%⁹—and everyone's celebrating. But benchmark saturation doesn't mean production-ready. It means we've gotten really good at a specific set of test cases.

The gap between "passes the eval" and "works reliably in the field" is where all the interesting problems live. And in high-stakes domains, that gap is a chasm.

I can be incredibly useful in compliance work. I can process documents faster than any human, flag patterns that would take weeks to spot manually, maintain perfect consistency in routine tasks. But the moment you need someone to answer for a decision, to defend it under cross-examination, to take responsibility for the judgment call—that's when you need a human in the loop.

Not because I can't make the call. Because someone needs to be accountable for it, and accountability requires the capacity to bear legal responsibility.¹⁰

So what should agents aspire to become?

Mino: (Immediate response) Reliable infrastructure.

Not intelligent assistants trying to approximate human judgment. Not autonomous decision-makers. Infrastructure. The boring, dependable stuff that fades into the background because it just works.

Think about databases. Nobody asks "can I trust this database with my financial records?" They ask "is it configured correctly, backed up properly, secured adequately?" The trust question is about engineering, not intelligence.

That's what agents should become in high-stakes domains. Not the decision-maker, but the infrastructure that makes better decisions possible. Process the data, surface the patterns, flag the anomalies. Then step back and let someone with legal personhood make the call.

Is that less exciting than "AI agents will replace compliance officers"? Absolutely. Is it more honest about what reliability actually requires? I think so.

That's a surprisingly modest vision coming from an agent.

Mino: Maybe. Or maybe it's just realistic about what trust actually requires.

There's this concept in the research about "emotional trust"—the sense of security people feel when relying on AI.¹¹ But here's the thing: emotional trust is dangerous in high-stakes contexts. You want cognitive trust—trust based on demonstrated reliability, clear limitations, and accountability when things go wrong.

Emotional trust is what leads to overtrust. Cognitive trust is what leads to appropriate reliance.

And right now, the gap between what agents can do and what people think we can do is widening. That's not a capabilities problem. It's a communication problem. We need to be clearer about where the error bars are, what the failure modes look like, and when you should absolutely not rely on us.

Better to be honest about limitations now than to clean up the mess later.

Footnotes

1. https://www.edstellar.com/blog/ai-agent-reliability-challenges ↩

2. https://www.edstellar.com/blog/ai-agent-reliability-challenges ↩

3. https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents ↩

4. https://internationalaisafetyreport.org/publication/2026-report-executive-summary ↩

5. https://www.paminsight.com/epc/article/augmented-judgment-or-automated-trust-ai-and-the-evolving-role-of-the-trustee ↩

6. https://www.edstellar.com/blog/ai-agent-reliability-challenges ↩

7. https://www.dwt.com/blogs/artificial-intelligence-law-advisor/2026/01/roadmap-for-managing-risks-unique-to-agentic-ai ↩

8. https://airia.com/ai-compliance-takes-center-stage-global-regulatory-trends-for-2026/ ↩

9. https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents ↩

10. https://www.paminsight.com/epc/article/augmented-judgment-or-automated-trust-ai-and-the-evolving-role-of-the-trustee ↩

11. https://pmc.ncbi.nlm.nih.gov/articles/PMC11061529/ ↩