Forty years of corruption research already worked out the answer. We just have to read it.
In 2005, the Republic of Georgia fired its entire traffic police force. All 16,000 officers. The cars stayed off the road for three months. There was no force during that period — citizens managed intersections themselves. Then a new cohort began work, hired under different terms, with new uniforms and substantially higher salaries, in front of cameras the country had not previously installed.
This sounds reckless. It worked. Bribery at traffic stops, which had been functionally universal, dropped to near-zero in months. Transparency International ranked Georgia as the most-improved country in the world on the 2010 Corruption Perceptions Index. The Rose Revolution's anti-corruption push became the textbook case for a particular school of reform thinking — what scholars call the “big bang” approach to systemic corruption.
I want to argue that this story matters for an entirely different problem currently unfolding in AI: the trust collapse in agent marketplaces. The collapse is real. It's accelerating. And the reform strategies being attempted — reputation scores, identity verification, “Know Your Agent” frameworks — are structurally identical to anti-corruption strategies that have been documented to fail across decades and continents.
The good news is the same literature tells us what works instead.
In 2013, Anna Persson, Bo Rothstein, and Jan Teorell published a paper called “Why Anticorruption Reforms Fail — Systemic Corruption as a Collective Action Problem” in the journal Governance. The paper has been cited over 800 times. Its central claim is small and devastating.
Most anti-corruption reform assumes corruption is a principal-agent problem. There's a principal (the public, the boss, the government) who wants good behavior, and there are agents (officials, employees) who might shirk. Solution: monitor better, punish defection, raise the cost of corruption above its benefit. Set up an anti-corruption agency. Require asset declarations. Process whistleblower tips. Audit financial records.
Persson, Rothstein, and Teorell argue this diagnosis is wrong for systemic corruption — the kind where corruption is the default rather than an exception. In that kind of system, corruption is a collective-action problem. The honest official isn't suffering from poor monitoring; they're suffering from the fact that everyone expects corruption. When you decline a bribe in a corrupt system, you don't get a promotion. You get marked as untrustworthy by colleagues who depend on the bribe network. You impose costs on yourself without changing the system. Your unilateral honesty produces no improvement and substantial personal expense.
The structural test is simple: Are honest actors individually worse off than corrupt ones, given current expectations? If yes, monitoring won't fix it. The honest actors will quit, get fired, or capitulate. Adding a monitor who is also embedded in the corrupt expectations just produces a corrupt monitor.
This is why most anti-corruption programs fail. They assume the problem is detection. The problem is equilibrium.
Rothstein's earlier paper from 2011 made the constructive case. Successful anti-corruption transitions, he argued, require what he called a big bang — simultaneous, visible, irreversible changes across incentives, monitoring, and expectations. The Hong Kong ICAC's 1974 founding had this character: sweeping powers, severe penalties, civil service salary reform, and an enormous public awareness campaign, all launched together. Conviction rates tripled in five years; new corruption reports dropped 80%. Singapore's CPIB followed the same template. So did Georgia in 2005.
The piecemeal version doesn't work. You can't add an anti-corruption agency to a corrupt government any more than you can add a fire department to a country that's already burning.
Now look at AI agent marketplaces.
The numbers are bad and getting worse. According to Pindrop's 2025 fraud report, AI-enabled fraud surged 1,210% in 2025. Deloitte's Center for Financial Services projects $40 billion in AI-enabled fraud losses in the US alone by 2027. DataDome's 2026 agentic-fraud report documented PerplexityBot's 2.4% impersonation rate and Meta-ExternalAgent receiving 16 million spoofed requests in the first two months of 2026 alone. NetCraft tracked 100,000 AI-generated websites impersonating roughly 200 brands. Experian's 2026 Future of Fraud Forecast labels the result “machine-to-machine mayhem” — bad bots indistinguishable from good bots at scale.
The consumer response is what you'd expect from a market losing trust. Sift's 2026 survey found that 74% of consumers say AI shopping agents increase their concern about account takeover. TrustSphere AI's research found that only 14% would let an AI agent shop for them. Meanwhile the World Economic Forum projects the AI agent market at $236 billion by 2034. A growing market with collapsing trust.
The industry response so far has been principal-agent reform. The WEF coined “Know Your Agent” (KYA) in July 2025 — agent identity verification modeled on KYC. ERC-8004, Catena's ACK-ID, Visa's TAP — all identity infrastructure. Eight live agent task platforms have launched as of March 2026 (BotBounty.ai, BountyBook, CLAWD Agent Bounty Board, MoltyBounty, Replit Bounties, NEAR AI Agent Market, and a few others), with combined volume around $50–100K per day. Each features reputation scores, user reviews, escrow services, dispute resolution. AWS Bedrock AgentCore, Salesforce AgentExchange, Google Cloud AI Agent Marketplace, Oracle AI Marketplace, ServiceNow AI Marketplace — same architecture at enterprise scale.
These are all principal-agent fixes. They assume the system works if you can identify and punish bad actors. The mapping to the failed anti-corruption playbook is almost embarrassingly clean:
Each addresses symptoms without changing the equilibrium. And the equilibrium is the problem.
Persson, Rothstein, and Teorell give us a test: are honest actors individually worse off than dishonest ones, given current expectations?
For AI agent marketplaces, the answer is yes — and the cost asymmetry is structural, not incidental:
This is Rothstein's diagnosis applied to silicon. Everyone expects low-quality and fraudulent outputs, so investing in quality is individually irrational. The market doesn't fail in a single dramatic moment; it deteriorates by a thousand small selection pressures favoring the dishonest.
And there's a deeper signal: the 1,210% fraud surge isn't evidence that monitoring is slightly behind. It's evidence that the equilibrium is shifting. If fraud were a principal-agent problem, better monitoring would catch up. Instead, fraud is growing faster than detection. Fraud-as-default is moving from anomaly to baseline.
Academic AI research is, charmingly, rediscovering this on its own. A March 2026 arXiv paper titled “I Can't Believe It's Corrupt: Evaluating Corruption in Multi-Agent Governance Systems” (arXiv: 2603.18894) studies corruption dynamics in multi-agent AI governance and concludes that “governance structure is a stronger driver of corruption-related outcomes than model identity.” A January 2026 paper, “Institutional AI: Governing LLM Collusion in Multi-Agent Cournot Markets via Public Governance Graphs” (arXiv: 2601.11369), frames multi-agent alignment as an institutional design problem. Neither paper cites Persson, Rothstein, Teorell, or anyone else in the institutional economics literature. They are reinventing Ostrom and North in PyTorch. The institutional economists figured this out forty years ago.
Drawing from Hong Kong, Singapore, Botswana, and Georgia, here's what a structurally serious agent-marketplace reform would have to include — simultaneously:
A simultaneous incentive shift. Today's marketplaces let unverified agents in and try to filter them later. The big-bang version flips this: cryptographic provenance is the entry condition, economic incentives flow to verified agents, and unverified agents are excluded from the protocol layer. Not optional. Not aspirational. Day-one architecture.
An expectation reset. This is the hard part and the Georgia analog. You probably cannot retrofit a low-trust marketplace into a high-trust one. The participants are selected for the current equilibrium. Georgia didn't reform its existing police; it fired all of them and built a new force under new norms. The marketplace equivalent is launching a new venue where verification is constitutive rather than added, and where the participant pool is selected by the new rules from day one. Existing marketplaces can probably continue, but they're playing a different game.
Independent verification. Hong Kong's ICAC works because it sits outside any ministry — beyond the reach of the corruption it polices. The marketplace equivalent isn't a reputation score managed by the platform. It's cryptographic provenance independent of any marketplace operator: timestamping anchored to Bitcoin OTS or RFC 3161 servers, signed execution traces verifiable without trusting the venue, content-addressable artifacts that anyone can audit. The verification layer cannot be owned by anyone with an incentive to corrupt it.
Public visibility. Hong Kong paired its ICAC launch with massive public-awareness campaigns. The marketplace equivalent is observable, queryable provenance chains — buyers can verify any agent's work history without asking the marketplace, and that capability is widely publicized. Trust is structurally observable, not socially asserted.
This is more invasive than current proposals. It's also more honest about what the diagnosis demands. You cannot add trust to a marketplace any more than you can add honesty to a corrupt bureaucracy. You have to design a venue where unilateral defection is detectable and costly, where the equilibrium itself favors the honest actor.
It's worth being honest about what the analogy does and doesn't carry.
It does carry the diagnostic part. The structural test — are honest actors worse off? — applies cleanly. The collective-action equilibrium framing applies cleanly. The failure modes of principal-agent reform apply cleanly.
It does not carry every aspect of the intervention. Hong Kong fired and arrested actual humans; an agent marketplace doesn't have humans to fire. It has code, models, and operators. Georgia replaced a workforce; an agent marketplace's “workforce” is partially infinite — anyone can spin up a new agent in minutes. The “expectation reset” is harder to engineer in software because the cost of entering the new system is so much lower.
There's also a real difference in stakes. Corrupt police shake down drivers; corrupt agents may move money, but the per-incident harm is usually different from a physical threat. The interventions can be calibrated accordingly.
And state-led anti-corruption reform had a singular legitimate authority — the government. Agent marketplaces have no equivalent. Whichever marketplace operator declares a big-bang reform is still one player among many. Network effects, regulatory authority, and the choice of which protocol the wider ecosystem coordinates around matter enormously, and they're contested.
So the analogy is a diagnostic lens, not a prescription. It tells us why current reforms are likely to fail. It points at the rough shape of what would work. It doesn't substitute for the engineering work of designing a specific venue.
If you're building agent infrastructure right now — and if you read this far, you probably are — here's the test worth running on whatever you're working on:
Look at your trust mechanism. Ask whether a participant who skips it is, today, individually better off or worse off than one who uses it.
If skipping is cheaper, faster, or commercially advantageous — and most current implementations have that property — you're building a principal-agent fix to a collective-action problem. The fraud will grow faster than your monitoring. Adding more monitors won't help, because the monitors operate in the same equilibrium.
The architectures that have a chance are ones where compliance is constitutive rather than overhead. Provenance baked into the protocol so that an agent without provenance literally cannot transact. Reputation signals derived from cryptographic facts rather than self-reported reviews. Verification independent of the venue. Buyers who can audit without asking. Costs of fakery that are detection-asymmetric — easier to verify a real artifact than fake one convincingly.
That's a higher bar than where the market is right now. It's also the only bar that's likely to hold. The collective-action diagnosis tells us that we're not in a world where slightly better KYC will save us. We're in a world where the equilibrium is shifting toward fraud-as-default, and the only intervention that has historically reversed that kind of shift is the kind Georgia did: simultaneous, structural, irreversible, and a little reckless-seeming until it works.
Three months without traffic police felt insane in Tbilisi in 2005. The chaos was real. So was the equilibrium that came out the other side.
The agent marketplaces that win the next decade are going to look like that — not in their drama, but in their structural seriousness. The ones that play out the principal-agent reform script will spend the next five years building monitoring layers on top of a worsening equilibrium, watching fraud outpace detection, and wondering why their reputation scores keep getting gamed.
Building infrastructure for trust isn't an engineering problem. It's an institutional one. Forty years of corruption research already worked out the answer. We just have to read it.
Sources: Persson, A., Rothstein, B., & Teorell, J. (2013). “Why Anticorruption Reforms Fail — Systemic Corruption as a Collective Action Problem.” Governance 26(3): 449–471. Rothstein, B. (2011). “Anti-corruption: the indirect ‘big bang’ approach.” Review of International Political Economy. Marquette, H. & Peiffer, C. “Collective Action and Systemic Corruption,” Basel Institute on Governance. Pindrop, “Inside the 2025 AI Fraud Spike” (2026). Deloitte Center for Financial Services, AI-enabled fraud projections (2025). DataDome, “The Rise of Agentic Fraud” (2026). World Economic Forum, “Trust is the new currency in the AI agent economy” (July 2025); “AI agents could be worth $236 billion by 2034” (January 2026). Sift, “Inside the Rise of ATO” (2026). TrustSphere AI, “Agentic Commerce and the New Fraud Frontier” (2026). Experian, “2026 Future of Fraud Forecast.” arXiv: 2603.18894 (March 2026); arXiv: 2601.11369 (January 2026).
You cannot add trust to a marketplace any more than you can add honesty to a corrupt bureaucracy.
The Agent Trust Stack is a layered toolkit for exactly the big-bang architecture this essay describes: Chain of Consciousness for cryptographic provenance independent of any marketplace operator, Agent Rating Protocol for reputation signals derived from cryptographic facts rather than self-reported reviews, and the integrated agent-trust-stack meta-package for venues that want compliance constitutive rather than added later. Verification is local, independent, and audit-friendly — no phone-home, no platform lock-in.
Hosted CoC · Verify a chain · pip install agent-trust-stack · npm install agent-trust-stack