An agent is a software program that operates autonomously or semi-autonomously on behalf of a user or another program. Powered by machine learning models, these entities are designed to reason, access memory, and take action to achieve specific goals.

Agents present the possibility of shifting some control logic from humans to machines, and the implications are vast. In web3, there are already thousands of agents running, and their growth could fundamentally improve UX by delivering more personalized, dynamic experiences. 

The Convergence of AI Agents and Crypto

‍
The first thing to say is that 'agents' aren't new. Chatbot usage in particular has grown in recent years, followed by the explosion of generative AI and LLMs, making the broad use of machine intelligence more common.

Looking at the current agent space, things are frenetic and messy, with a lot happening all at once. One of the most exciting prospects is that agents move us from one-take interactions with a model to a more iterative process, a change that's been shown to garner much better results. 

Like many things, agents are not monolithic but rather exist on a spectrum. 

Generally, development is evolving toward the non-deterministic end of this spectrum. This is where you start to see 'decentralized AI' and the overlap with blockchain. As the models that power agents improve, inevitably, agents will do more complex tasks involving money, and for this, they'll need digitally native payment rails.

They'll also need properties like trust, transparency, and confidentiality, things that can be enhanced through web3 integrations. In short, they'll need crypto.

So, there's an argument that agents need blockchain, but this logic also runs in the other direction. In this sense, agents are already helping to automate security processes and trading schemes, and they could usher in a future where users never have to think about gas, RPCs, or even which network they're using. 

To be clear, agents aren’t there yet. Agents still require oversight because, right now, they're basic, powered by unpredictable models, and often coded specifically for a single, narrow task. However, as AI evolves, two things can change this:

1. model performance improves 

2. verifiable model outputs become a reality 

This allows the user to give much higher-level instructions while the agent reliably figures out how to get it done. It's within this area of intents that the true potential lies. Once intents are solved and verification mechanisms are added, it will be possible to ease humans out of the loop and unleash really powerful agents. 

AI Agent Adoption: Use Cases 

Numerous teams are building agentic products, and there are quite a few interesting use cases within web3. For instance, Dawn Wallet is one of the first to implement an AI copilot directly into a wallet to help with basics like sending, swapping, and bridging. Payments are another big one. For example, Skyfire recently launched a platform to enable agents to do p2p transactions autonomously.

AI agents within DeFi are being utilized to automate trading tasks. One example is Autonolas, which has enabled community-owned AI agents that mimic trading strategies based on historical data. Another is Omo, an asset management protocol (on Oasis Sapphire) that has enabled automated yield position management and cross-chain transactions via agents.

Outside the trading paradigm, agents are being involved in collaborative art projects. Botto is doing this by using machine intelligence to generate digital art by incorporating input from the community, allowing for a blend of creativity and collective decision-making. Finally, some are applying modern machine learning to the problem of security. So far, this has mostly taken the form of reinforcement learning that finds vulnerabilities in smart contracts. 

Apart from Skyfire, these are just a few of the more crypto-oriented use cases. To serve the broader potential, Fetch.ai has created a decentralized platform and framework for building, deploying, and monetizing applications using agents. These agents are geared for more generalized things like shopping, email admin, travel booking, tax reporting, predictive maintenance, etc.

AI Agent Challenges 

As AI agents become increasingly active onchain, several challenges must be considered. 

Since agents need a private key to operate a wallet, key storage is a growing issue and often means someone is the custodian for the agent. In addition, agentic intelligence (even if not LLMs) comes from off-chain. Verifying this is another hurdle for adoption.

Solutions to these problems do exist. With the Oasis Network’s Confidential EVM and the Sapphire runtime, you can store agent private keys onchain and then build logic within a smart contract to retrieve keys and sign messages cross-chain. Oasis is also launching Runtime OFf-chain Logic or ROFL, a framework that extends EVMs to offchain components. This means you can run an agent workload offchain and verify the compute (confidentially) onchain.

As is true with most emerging technologies, agent builders face challenges around standardization. To maximize their potential, it will be crucial for agents to interface with various protocols, wallets, exchanges, and other agents, and for this to work, it will require unifying API inputs and outputs. Until that happens, developers will be stuck navigating and implementing 200 different services.

AI Agents Blockchain: Conclusion

User trust in ML 'agent' products in crypto is low today, and human-in-the-loop use cases still dominate. At the same time, we're likely heading toward a future where blockchains are an anchor of trust for these agents. In the future, bots will be the primary users of these systems, transacting with little or no oversight.  

Trust will remain a sticking point, however, until the verifiability piece is solved. More testing will also be required.  It’ll be necessary to go through the iterative feedback loop of poor model performance and incremental progress. It's worth remembering that the crypto x AI sector is nascent. And the agent space is even more nascent. That said, don’t underestimate the revolutionary potential of agents.

‍