Complexity Abstraction: AI Agents and the emerging web3 UX
![Complexity Abstraction: AI Agents and the emerging web3 UX](https://cdn.prod.website-files.com/63739957c8b86d3a88ae7085/668fbc0595da122140a6b0f2_Biconomy_AIagent_10Jul.webp)
AI x crypto is an emerging field that will unlock tons of powerful use cases. However, one key area that is currently under appreciated is how AI will take web3 UX to the next level. We believe it will play a big role in abstracting certain complexities to simplify how users interact with web3 in the future.
Here's a deep dive into how this emerging AI agent powered UX may look like.
The current state of Web3 UX
Web3 UX is a tough problem to solve. Since 2018 we have seen very real progress when it comes to improving the UX of a specific Dapp thanks to account abstraction, app chains, social logins, passkeys, and gas relayers to name a few.
Make no mistake, these technologies have not finished their progression. Account abstraction is getting better with big releases this year and more players entering the game. Chain abstraction and intents will certainly help navigate the UX hell of bridging…but there is a web3 UX developing which leverages all of the work done since 2018 and pushes the UX status quo to a place that surpasses web2 UX.
The complexity translator
Whenever a technology is invented which does something thought to be impossible, it will have a large impact on society. This is true of Bitcoin and is true of LLMs. LLM’s superpower is that it can transform content into a new form.
Web3 transactions are code, instructions to the blockchain in a language it natively understands, 1s and 0s. Humans, with the exception of web3 devs and protocol developers, for the most part do not understand code. And this is where LLM’s are so powerful from a UX perspective. They can translate human language to code and vis versa, removing the need to understand code and working with you on terms you understand.
Personalised Technical Expertise
What this going to feel like is having your own seasoned smart contract tech lead by your side when using web3. This experience of working with an expert who can abstract away low level complex concepts like addresses, nonces, call data will make transacting on web3 feel smooth, intuitive and creative as mainstream users become unburdened by complexity. LLM’s as a foundation for mainstream interaction with web3 is going to take web3 UX much further than just been on par with web2, it's going to change it completely.
AgentFI, the dawn of a UX revolution
Leveraging LLMs to abstract away user complexity is happening right now by a number of different players. This section aims to breakdown the UX and its application.
Expert Planning
The interaction with an AI agents is not about the AI devising plans on its own but rather facilitating a dialogue guided by user curiosity. This conversation generates up-to-date, well-informed, and accurate data points and counterpoints, akin to having your own personal web3 expert working alongside you. The end result of the dialogue is a game plan.
Kaito AI is a promising web3 research agent which embodies this UX. It takes a vast amount of crypto market data and allows you to engage with it on your terms.
![](https://cdn.prod.website-files.com/63739957c8b86d3a88ae7085/668fbaad2ad84a68193305d9_Screenshot%25202024-07-05%2520at%252015.38.03.webp)
Where will this go?
Specialisation and pay per prompt. This is my personal take, but I think research agents are going to specialise based on their access to data, model strategy and the feedback they get back from their users. I expect there to be a “General” planning agent which will reach out to these specialised agents for specific questions and then pay for it. Expect to see a model of “Pay per Prompt”, and luckily crypto payment rails makes such transactions trivial. This is not a small point, in-fact programable money is why we are going to see cutting edge UX emerge on crypto rather than on web2.
![](https://cdn.prod.website-files.com/63739957c8b86d3a88ae7085/668fbae5760d39deaf788678_Screenshot%202024-07-11%20at%202.56.11%20PM.webp)
Expert Execution
Planning is good, but the AI agent is going to have to make transactions on the user’s behalf if its going to change the UX. The transactions it executes are not going to be simple ones that could be facilitated by existing UIs. Transaction execution facilitated by agents will need to excel along four dimensions:
- Precision: The agent must operate with flawless accuracy due to the high financial stakes involved. We have actually seen elements of this with Metamask’s Blockaid integration and Safe’s tenderly simulations, two useful tools to help users execute transactions which are aligned to their intentions.
- Transaction Efficiency: Beyond executing single transactions, the real value of AI agents lies in their ability to batch multiple transactions efficiently, saving users time and money. Furthermore the selection of protocols will be important as agents balance transaction costs and security when selecting a protocol to work with.
- Multi-Step Operations: The agent should handle multi-step processes, such as bridging assets across chains, involving several non atomic transactions.
- Autonomy: For time sensitive tasks, the agent should autonomously take action. This can be as simple to wait for assets to be bridged to more complex actions like claiming an airdrop.
Where will this go?
Beyond clever architecture to eliminate hallucinations and a robust knowledge graph mapping user’s plans to transactions I think there are few ways accuracy could be scaled for initial agents.
- Permission Packs. Much like Uniswap's token list, provide agents the ability to leverage a pack of permissions which relate to an area of web3. We could image an DeFi pack which gives the AI permission to interact with known defi addresses and functions. In this way removes the risk of interacting with a hallucinated addresses or functions.
- Transaction Insurance. A potential new business model, the execution agent can take a fee and insure the transaction will execute on the users goal. If it does not it loses some stake or value.
This section requires its own post to really dive into the complexity of executing transactions flawlessly and the emerging strategies to tackle it.
Where is the application?
Broadly speaking we can see AI entering web3 UX in two ways, general and specific.
Specific
Specific application of AI agents will permit a limited set of interactions on specific protocols. Because the complexity is narrower than a general application of AI powered web3 UX we can see it been adopted first.
A great example of this approach is Agent Coins (Prev Polywrap) fundpublicgoods.ai. The product simply asks the user what they would like to donate to on Gitcoin. It then leverages the researcher aspect to review all gitcoin grant projects, create weighted short list of ones which align with the user intention.
![](https://cdn.prod.website-files.com/63739957c8b86d3a88ae7085/668fbb2652f239fec4ba846a_Screenshot%25202024-07-05%2520at%252016.41.03.webp)
Then the execution aspect kicks in when the product prepares a batch transaction donating to the projects based on the weighting suggested by the AI. Although this is early, expect to see this UX proliferate across Dapps and DAOs in the coming months.
General
General Agents aim to take any request from the user and turn it into transactions. These agents have the potential to be the most transformative but they are also the most difficult to pull off given they have to handle any input the user may throw at them. There are no Agents in production that execute on this UX, although the demos Wayfinder.ai have been generating are the most compelling.
The general application will have massive knock on effects with respect to UX. Firstly its effect on personalisation with be enormous. With AI’s been able to mould the experience so closely to the individual that uses it, “not understand how it works” will become a thing of the past as the experiences starts to feel like an extension of ones self.
The surface has not been scratched
At a high level complexity will be abstracted away from end users but there is so much more. The tech stack that will power this UX is interesting and complicated all the way form the smart contract wallets to how compute is routed to GPUs to process the prompt. Privacy is going to be crucial and so will ensuring the outputs from the models are actually from the models you think you are interacting with. Decentralisation will also be extremely important in terms of which parts of the AI stack require it over other and how decentralised AI can compete with the rapid changes in AI models. I will be writing on these in the coming weeks but the main take away is this is coming and Biconomy will be doing our part in bring this UX to the mainstream.
------------------------------------
This piece was authored by Graeme Barnes