The ledger remembers what the mempool forgets. On March 12, 2027, ChainCowork—Anthropic’s blockchain-focused AI subsidiary—dropped a feature that the automation industry didn’t see coming: Record Workflow. It lets users record screen interactions, mouse clicks, keystrokes, and voice commands, and instantly compiles them into a reusable on-chain skill. Three days later, BlockCodex, a rival from OpenAI, released an identical feature under the same name. The market cheered. I did not.
I spent the last 72 hours parsing the API logs, contract bytecode, and user session data from both platforms. What I found is not innovation. It is engineering convenience dressed as a breakthrough. Both teams took existing multimodal models—Claude 3.5 and GPT-4o—and wrapped them in a screen-recording pipeline. The output is not a smart contract; it is a structured prompt that references an SKILL.md file containing natural language steps and script templates. This is behavioral cloning applied to browser automation. Code is not law, it is merely preference. And here, the preference is to pretend that recording equals creation.
Context: The hype cycle that forgot the blockchain The blockchain industry has spent 2026 chasing AI-agent narratives. Projects like Fetch.ai, Render, and Bittensor have seen 300% valuation pumps on promises of autonomous agents managing DAO treasuries, executing trades, and auditing code. Yet the actual infrastructure remains fragmented. Most “AI agents” are simple GPT wrappers that call an API. ChainCowork and BlockCodex are the first to package this into a desktop application that claims to generate on-chain actions—such as submitting transactions, interacting with DeFi protocols, or compiling deployment scripts—by simply watching a user perform the task once.
But the gas wars expose the cost of decentralization. The recorded skill is stored off-chain, in Anthropic’s or OpenAI’s cloud. Only the execution trigger is on-chain—a hash pointing to the skill ID. This centralizes the automation logic, turning the blockchain into a mere logging layer. The illusion persists until the liquidity dries. Right now, the liquidity of trust is high. Users believe they are creating verifiable, immutable automations. They are not.
Core: A systematic teardown of the Record Workflow architecture I reverse-engineered the recording pipeline using a clean Ubuntu instance, a hardware keylogger for validation, and a microphone to capture voice commands. Here is the raw data flow:
- Phase 1 – Capture: The ChainCowork client records screen area (configurable), mouse coordinates (x,y every 50ms), keyboard inputs (scancodes with timestamps), and audio via the user’s mic. This data is streamed to Anthropic’s API gateway. No local processing. The packet size averages 4.2 MB per minute of recording.
- Phase 2 – Parse: The multimodal model (Claude 3.5 Sonnet) processes the video frames, extracts OCR text, identifies UI elements (buttons, fields), and transcribes the audio via a Whisper variant. The output is a JSON sequence of actions:
[{"action":"click","element":"submit-button","text":"Submit"},{"action":"keypress","value":"password123"}]. Notice: no semantic context, just raw imitation. - Phase 3 – Compile: The action sequence is inserted into a pre-defined SKILL.md template. An auxiliary code generator (likely Codex or a fine-tuned model) converts the sequence into a Python script that uses Selenium or PyAutoGUI. The script is stored on Anthropic’s server. The user receives a skill ID and a one-click execution button.
- Phase 4 – Execute: When triggered, the client downloads the script and runs it locally. The execution environment is sandboxed—no internet access except to the target blockchain RPC. But the sandbox is user-configured. I found that 30% of testers disabled sandboxing for “performance reasons,” exposing their entire system to the recorded actions.
Floor prices are just liquidated confidence. The execution failure rate for my 50 test skills (ranging from token swaps to NFT minting) was 38%. Common failure modes: UI layout changes (24%), authentication token expiry (9%), and model hallucination generating incorrect click coordinates (5%). The system does not retry; it aborts with an error message. A skill that works today will break tomorrow.
Contrarian: What the bulls got right I must acknowledge the counter-intuitive strength. The low-code automation market for blockchain workflows is real. Current RPA tools like UiPath require a dedicated developer to script interactions with MetaMask or MyEtherWallet. ChainCowork and BlockCodex drop that requirement to zero. For simple, repetitive tasks—daily claim of staking rewards, logging into a governance portal, signing a batch of transactions—the record-and-replay approach is sufficient. In my tests, a skill to claim Yield Farming rewards on Polygon took 2 minutes to record and executed successfully 9 out of 10 times. The tenth failure was due to network latency.
Moreover, the competitive dynamic is forcing both teams to improve execution reliability. BlockCodex already introduced a “skill refresh” feature that automatically retries with adjusted selectors. ChainCowork added a dry-run mode that simulates execution against a forked mainnet. These are genuine enhancements that benefit all users. The network effect of shared skill libraries—should they open them—could create a formidable ecosystem. Truth is a derivative of transparent data. And for now, the data shows that the low-end use cases work.
Takeaway: The accountability call The biggest risk is not technical failure. It is the illusion of immutability. Users will build long-running automation workflows—treasury management, automated trading strategies, compliance reporting—on top of these recorded skills, believing they are “on-chain” and therefore permanent. They are not. The skill definition lives inside a centralized API. If Anthropic or OpenAI shuts down, changes pricing, or updates the SKILL.md format, every skill breaks. We debugged the narrative, not the contract.
Immutability is a feature, not a virtue. The virtue here is convenience. But convenience without auditability is a liability. My recommendation: treat Record Workflow skills as temporary macros, not permanent smart contracts. Test them weekly. Never store private keys in the recording. And demand that both providers publish the exact schema of SKILL.md along with a hash-anchored versioning on Ethereum or Arweave. Until they do, the ledger remembers what the mempool forgets: these tools are prototypes, not pillars.
