The launch of Claude’s "Record a skill" feature—and its near-identical twin from OpenAI’s Codex—isn’t just a product update. It’s a macro signal. Two of the most capitalised AI labs have converged on the same engineering thesis: the next frontier of value capture is automating human GUI behaviour. For crypto, this is both a liquidity accelerant and a systemic risk multiplier.
Hook
On February 15, Anthropic quietly enabled Pro, Max and Team subscribers to record their screen, clicks, keyboard input and voice, then package that demonstration into a reusable "skill". Eleven days earlier, OpenAI had added an identically named feature to Codex. The margin of differentiation is zero. This is not a coincidence. It is a tactical alignment on a shared technical conclusion: the highest-ROI application of current multimodal models is behavioural cloning applied to desktop automation.
Context: The Macro Map
Barcelona, 07:14. I’m staring at the global liquidity composite. US money market funds are shrinking. DeFi stablecoin supply is flat. Yet the two largest AI firms are pouring inference compute into a feature that lets a non-technical user teach an agent to fill a web form, copy a row from Excel to Salesforce, and send a Slack confirmation—all by simply doing it once. This is not a model architecture breakthrough. It is a product wrapper around existing capabilities: screen capture, ASR, intent parsing, and code execution. The technology is mature enough to be commoditised.
From my 2020 DeFi stress test work, I learned that leverage migrates to the point of least friction. In 2021, that was NFT wash-trading. In 2024, it was ETF inflows. Now, the friction is user interface. If you can record a process and delegate it to an AI agent, you have effectively turned any knowledge worker into a potential automation creator. That is a massive unlock of latent productivity. But in crypto, we don’t confuse volume with value. We ask: where does the counterparty risk sit?
Core: The Institutional Convergence Play
The core insight is simple: AI agents that automate UI tasks will drastically reduce the cost of operating complex workflows, including crypto-native ones. A compliance officer at a Barcelona family office can now record the process of transferring stablecoins from a CeFi account to a DeFi protocol—KYC checks, withdrawal triggers, contract interaction—and reuse that skill weekly. The agent executes the steps exactly as recorded, but it also learns from each run. The execution cost approaches zero. The business logic becomes a tradable asset.
History rhymes. This isn’t the first time we’ve seen this pattern. In 2021, I published "The Illusion of Scarcity" after tracking $50M in wash-trading volume across NFTs. The thesis was that retail FOMO masked a lack of genuine institutional interest. Today, the same dynamic applies to "recording skills". The feature will attract a wave of non-developer creators who will build thousands of automation scripts. The surface area for error—and for malicious skills—explodes. But the market will reward reliability, not novelty.
From a macro perspective, this accelerates the institutional convergence I quantified in 2024 when I argued that $40B in spot Bitcoin ETF inflows would flatten volatility and increase correlation with S&P 500 liquidity cycles. AI-driven automation does the same for operational cost curves. It compresses time. A task that took a junior analyst two hours now takes thirty seconds. That efficiency feeds directly into portfolio churn rates, transaction velocity, and ultimately, on-chain liquidity.
Contrarian: The Decoupling Thesis is a Mirage
The prevailing narrative is that AI automation decouples crypto from traditional macro cycles. I reject that. The opposite is true. When both Anthropic and OpenAI centralise skill storage on their own servers, they reintroduce the exact counterparty risk that crypto was designed to eliminate. A recorded skill contains hardcoded API keys, wallet addresses, and private workflow logic. If that skill is shared—and a marketplace is inevitable—the data ownership and execution integrity become single points of failure. This is not decentralisation. It’s rebundling of trust into a few proprietary clouds.
In the bear market of 2022, I learned that counterparty risk is the primary macro driver. Celsius, BlockFi, Terra—all failed because trust was centralised. Today’s "skill market" carries the same DNA. The skills themselves are opaque black boxes. There is no on-chain verification of what a skill actually does. A malicious skill could exfiltrate a multisig key or submit a fraudulent transaction while appearing to perform a legitimate action. The forensic liquidity skeptic in me sees the parallel: just as exchange "Proof of Reserves" audits were theatre, so too will be the "skill audit" processes of these AI platforms, unless they implement continuous, verifiable execution traces.
Furthermore, the feature’s reliance on screen recording and keylogging introduces a privacy vector that enterprise compliance teams will struggle to accept. I’ve seen this before: in 2017, when I wrote the scalability trilemma whitepaper, the bottleneck was technical. Today, the bottleneck is institutional trust. No regulated financial entity will allow its employees to record every mouse click and password entry and upload that data to Anthropic’s inference endpoints. The hidden assumption in the marketing copy is that users will trust the platform. That assumption breaks against the reality of SOC2 audits and GDPR fines.
Takeaway: Cycle Positioning
The question every macro strategy analyst should ask is not whether AI agents can automate workflows. They can. The question is: who captures the economic surplus? My positioning is to short the platforms that centralise skill execution and long the protocols that enable verifiable, permissionless automation. Code doesn’t confuse volume with value. It executes exactly what it’s told. The next cycle’s alpha will come from identifying the infrastructure layer that bridges AI agents to on-chain execution without reintroducing counterparty risk. That layer doesn’t exist yet. But the macro signal is clear: the race to record the playbook has begun. Follow the capital flows, not the feature announcements.
