A few weeks ago, I watched a product manager record a 90-second screen session in Claude, and the system turned it into a reusable "skill." No code, no drag-and-drop โ just a demonstration. By the end of the week, her team had automated their entire monthly reporting process. Meanwhile, across the aisle, OpenAI's Codex quietly launched the same feature. Two AI giants, same feature, same week.
But here's the twist: neither of these skills lives on a blockchain. They live on proprietary servers, owned by centralized entities. As an open-source evangelist who cut my teeth auditing tokenomics in Hangzhou's 2017 ICO frenzy, I couldn't help but ask: what happens when the next generation of automation โ the very workflows that will shape how we work, trade, and govern โ is locked inside corporate walled gardens?
This is not just a product update. It is a fork in the road for Web3's promise of self-sovereign automation. And it's a battle that will define whether the future of work runs on decentralized protocols or on rented APIs.
Context: The Rise of Record-and-Repeat Agents
Both Anthropic and OpenAI have converged on a similar technical pattern: use a multimodal model to record screen activity, clicks, keyboard inputs, and voice, then compile that demonstration into a reusable script or prompt that can be replayed later. The underlying mechanism is a flavor of behavior cloning โ an imitation learning technique where the agent learns a policy for performing a task in a GUI environment by watching a human do it once.
From a blockchain perspective, this is the equivalent of a smart contract that is written not by a developer, but by a user demonstrating the desired behavior. The implications are enormous for DAOs, DeFi, and any on-chain workflow that involves off-chain components. Imagine recording a multisig signing ceremony once, then having an agent replay it with different parameters every month. Or recording the steps to claim an airdrop across multiple chains, then sharing that skill with your community.
But here's the catch: today's implementations are completely centralized. The skills are stored on Anthropic's or OpenAI's servers, executed by their models, and governed by their terms of service. Your automation workflow is not yours โ it's theirs.
Core: The Decentralization Gap
Based on my experience auditing token economics and community governance proposals, I see three layers where this centralization creates risk.
First, execution relies on a single API provider. If OpenAI goes down, your accounting skill fails. If Anthropic changes pricing, your cost per skill triples overnight. In Web3, we design systems where no single point of failure controls the execution path. A decentralized alternative would store skills on IPFS, verify execution proofs on-chain, and allow any compatible model to replay them.
Second, the training data pipeline is opaque. Every time you record a skill, you're sending your entire workflow โ including sensitive data like account numbers, internal URLs, and meeting discussions โ to a central server. The terms of service may allow the provider to use that data for model training. In a decentralized setup, you control the data: encryption before sharing, zero-knowledge proofs for verification, and local execution for privacy.
Third, skill composability is limited. Today, you can't chain a Claude skill with a Codex skill because they use incompatible formats and execution environments. On a blockchain, we have standardized the composability of smart contracts via interfaces and standards like ERC-20. The same should apply to agent skills: an open standard for skill definitions that allows any agent to execute any skill, regardless of which model generated it. This is exactly the kind of interoperability that Ethereum brought to tokens, and that we now need for automation.
Contrarian: Maybe Centralized Agents Are Actually Fine (For Now)
Let me play devil's advocate. A product manager I spoke with last week said, "I don't care if it's decentralized โ I just want it to work." And she's right. The current centralized agents work well because they are tightly integrated with powerful models and optimized for low latency. A decentralized alternative would need to solve real-time inference on decentralized compute, which is orders of magnitude more complex.
Moreover, the network effects of a shared skill marketplace could be massive. If Anthropic or OpenAI hosts the best repository of high-quality skills, users will flock there. The convenience of a one-click solution often outweighs the philosophical purity of decentralization, especially for non-technical users.
But here's the blind spot: centralized agents are the perfect surveillance tool. Every workflow you record becomes visible to the provider. Every automation pattern reveals your business processes, your reporting cadence, your compliance gaps. For enterprises that deal with regulated data (financial, medical, legal), this is a non-starter. The European Union's AI Act already classifies high-risk AI systems, and a skill that handles sensitive data may fall under that umbrella. The only way to guarantee compliance is to keep the data and execution within your own infrastructure โ which requires a decentralized, self-hosted option.
Takeaway: We Need an Open Standard for Agent Skills
The AI giants are racing to own the user's automation layer. If they succeed, the metaverse of workflows will be built on rented land. As a community, we should start defining an open spec for agent skills โ a format that is model-agnostic, verifiable on-chain, and privacy-preserving. Think of it as ERC-721 for automation behaviors. We need a governance framework for skill standards, a marketplace where creators can mint and sell their skills as NFTs with royalty guarantees, and a runtime that can execute on decentralized compute (think Akash or Golem) without ever touching a proprietary API.
The next six months will determine whether the future of work is interoperable or siloed. Let's not let it be decided by the same companies that brought us walled gardens in the first place. Trust isn't compiled in a cloud โ it's distributed, verified, and shared.