When the State Deploys an AI Agent: Why We Need Blockchain Governance for Artificial Intelligence
Every system that centralizes decision-making creates a single point of capture. Last week, a news flash from Beating AI News reported that Tencent is piloting WorkBuddy, an AI agent for government employees in Guangdong province. The agent automates policy checks, document drafting, and business system interactions. It promises efficiency, but as a cryptographer who has spent years designing decentralized governance, I see a pattern repeating: the tools we build shape who controls the future. WorkBuddy is not a breakthrough in AI; it is a breakthrough in centralized control. And that is precisely why the blockchain community needs to pay attention.
WorkBuddy is an application-layer product, not a new foundation model. It combines retrieval-augmented generation (RAG), tool calling, process automation, and local deployment. The agent can batch review maternity subsidy applications, draft policy materials, and write results back into government systems after human confirmation. The technical complexity lies not in the model but in the integration layer: unifying identity management, permission middlewares, API gateways, and audit logs. This is what I call a combinatorial innovation, not an architectural breakthrough. Yet the implications are profound. The agent sits inside the government network, accesses the same databases as civil servants, and executes actions within their permissions. It is a digital employee that never sleeps, never forgets, and never questions the rules it follows.
From a blockchain perspective, WorkBuddy is a centralized oracle with executive power. In decentralized finance, oracles provide data to smart contracts. Here, the AI agent is both the oracle and the executor. It reads government knowledge bases, interprets policies, and writes back to production systems. There is no public ledger, no cryptographic proof of correct execution, no mechanism for citizens to verify that the agent followed the rules. The government trusts Tencent's code, and the citizens must trust the government. This is a trust model that our industry has been trying to dismantle for over a decade. Based on my experience auditing decentralized protocols, I have seen how even well-intentioned centralized systems become opaque and prone to capture over time. The same will happen here unless we embed transparency by design.
The core engineering challenge of WorkBuddy mirrors the challenges of DAO governance: how to enforce permissions, how to audit actions, and how to ensure that the system behaves as intended. The article notes that the agent uses only the permissions that the civil servant already has. This is a sensible design, but it creates a blind spot. Who controls the permission system? Who updates the knowledge base? Who decides which policies the agent interprets? These are governance questions, not technical ones. The phrase "govern the entrance, not the exit" comes to mind. WorkBuddy's entrance is the permission system; the exit is the agent's actions. If we only govern the entrance, we miss the critical point: the agent itself is a black box that can be updated, fine-tuned, or replaced without public scrutiny. In a decentralized context, we would require a governance token or a multisig to approve such changes. Here, the decision is made behind closed doors.
Now, let us consider the contrarian angle. Centralized AI agents are efficient. They can be deployed quickly, integrated with legacy systems, and scaled with government budgets. The pilot in Guangdong will likely show impressive productivity gains. The article's analysis estimates that simple material review tasks could see 20–40% substitution within two years. This is good for the government's bottom line. But the real cost is not measured in dollars; it is measured in agency. When the agent becomes the primary interface between citizens and the state, the citizens lose the ability to understand and challenge the decisions that affect them. The blockchain community often talks about "code is law, but people are the soul." WorkBuddy is code that enforces law, but the people are not in the loop. The human-in-the-loop design only applies to the final confirmation step. The agent's reasoning, its policy interpretation, its handling of edge cases remain invisible. This is a recipe for erosion of public trust.
Furthermore, the article highlights that WorkBuddy is a B2G project with private deployment. This means the government pays for the software, the hardware, and the ongoing maintenance. Tencent gains a sticky customer locked into its ecosystem. From a market perspective, this is a sound business model. From a decentralization perspective, it is a nightmare. The government becomes dependent on a single vendor for critical infrastructure. In the event of a security breach, a licensing dispute, or a change in corporate strategy, the entire system could be compromised. The blockchain ethos suggests that we should distribute trust, not concentrate it. The same logic applies to AI governance. We need systems where the agent's actions are auditable by independent parties, where the rules are encoded in transparent smart contracts, and where any citizen can verify the agent's behavior without relying on a single authority.
My experience with DAO governance has taught me that the most resilient systems are those that anticipate failure. During the 2022 bear market, I saw dozens of centralized lending protocols collapse because their governance was opaque. The same pattern will emerge in government AI if we do not act now. The technology is not the problem; the governance model is. WorkBuddy uses a combination of OCR, RAG, and RPA, which are mature technologies. What is missing is a layer of cryptographic accountability. For example, the agent could produce a zero-knowledge proof that it correctly applied a policy to a given set of inputs. This proof could be published on a public blockchain, allowing anyone to verify the agent's compliance without revealing sensitive data. Such a system would preserve privacy while ensuring transparency. It would shift the trust model from "trust Tencent and the government" to "trust the math."
Of course, implementing such a system is not trivial. The government would need to standardize its policies into machine-readable formats, the AI agent would need to generate proofs of correct execution, and the blockchain would need to handle the throughput of millions of government transactions. Post-Dencun, blob data will be saturated within two years, and rollup gas fees will double again. But that is a scalability challenge, not a fundamental barrier. We can start with something simple: a hash of each agent action logged on a public chain, with a commitment to the policy version used. This would allow retrospective audits without real-time verification. The infrastructure exists today.
The takeaway is not that WorkBuddy is bad. It is a well-engineered product that will improve government efficiency. The takeaway is that we, as a community, must engage with this trend before it solidifies into a centralized architecture that is hard to change. The blockchain community has a unique perspective on trust, transparency, and governance. We should apply it to the AI agents that are increasingly running our societies. Let us not just build decentralized finance; let us build decentralized governance for artificial intelligence. The future is not just AI agents; it is AI agents that are accountable through verifiable, transparent, and distributed systems. Code is law, but people are the soul. And the soul of governance requires that every citizen can see the law in action.