This week we introduce two new features built for Tx Explain; both demonstrating composability of the agent's original output -- a human-readable transaction summary.
Quick links for Tx Explain:
Today's Release Notes
- Transaction Categorization (live now on the app)
- MetaMask Snap Development Preview (open-source repo on GitHub)
Transaction Categorization
For Agents to act autonomously, we need them to better understand their actions.
Transaction categorization works by taking the human-readable transaction summary produced from Claude and asking another LLM (in this case, Llama3) to categorize the transaction from a selection of predefined categories.
Possible categories: DeFi, ETH transfers, ERC-20 token transfers, L2, MEV, NFTs, New contract deployment, Blob transaction, Misc
The output is a scoring of the transaction's most likely categories as determined by the model.
Example output:
"labels": ["DeFi", "ERC-20 token transfers"], "probabilities": [0.8, 0.2]
This feature demonstrates that LLMs can decode a transaction into an accurate summary and use that output in a useful, composable way with other LLMs.
At scale, the automation of categorized transaction summaries unlocks possibilities beyond the category tags themselves. Automated categorization could enable:
- Cluster analysis and search for similar transactions
- Identify emergent trends (based on not fully matching any of the existing categories)
- Routing different transaction types to be handled by different services
MetaMask Snap Development Preview
In the previous release notes, we introduced support for unsigned transactions. Today, we open-source a fully functional MetaMask Snap, which presents a simple summary during the signing process.
Feedback
Please use the in-built feedback form to let us know about any features you'd like to see, or general comments you have about Tx Explain.