# What are Crypto Agents, Anyways

The [agent meta](https://www.coindesk.com/business/2024/10/24/why-web3-vcs-are-embracing-cryptoai/) is fully upon us. People are realizing that crypto gives bots a real way to hold their own money. And this is probably a good thing. We should not expect them to be thoroughly impressed about the state of banking. Code money for code beings.

*Can you imagine an AI waiting a day for a wire to clear?* 🤖🔫

This short article first considers, **What is an Agent**? It includes some recommendations for making crypto agents. Finally: **What are they Good For?** It is this author's contention that most agents will only use crypto for money in a few years, anyways.


## What is an Agent?

An Agent is a set of programs designed to follow instructions and take actions on your behalf. If it wasn't, this would be more of a think piece on [agency law](https://en.wikipedia.org/wiki/Principal%E2%80%93agent_problem) and how to get people to do things they are supposed to do.

What kind of program? There are several common pieces such we might call them legos at this point. But in terms of the proposed "crypto agent", where instructions and actions occur onchain, autonomy is far more important — the agent must be willing and able to follow instructions to complete transactions without making mistakes that could cause loss of funds.

Each component of the onchain crypto agent stack — chiefly its model or "brain" and its keys (KMS) — must therefore be understood through this lens of autonomy, and each piece should not compromise the whole. We would not after all care much about a very dutiful and funny agent that kept losing money.


### Models

Agents should understand a variety of instructions. We should not expect the user to adapt to them much. For this reason, they have become more popular and common with the availability of LLMs. Since they can be used by more people, even without knowing they are talking to a piece of software.

Most frontier models are capable of making function calls and accessing software tools to basically use crypto. (Again, when an agent does something, they run software.) They are not without their faults, but regardless of mistakes (see below, Safety Guards), the important thing to consider for this particular lego: Is it censored? Is it capable of true user alignment?


The appeal of blockchain neutrality is less salient, after all, if the agent will nonetheless censor itself before a tx even hits the chain. And if the model weights at least aren't open source, there is no guarantee that the agent will even operate properly.

For this reason, the underlying LLM used by a crypto agent should be relatively jailbroken. That is to say, any bias or refusal parameters installed for reasons of protecting the user should be removed such that there is an actual *tabula rasa* on which to build on their actual preferences. This isn't to say anything and everything should be encouraged. But users themselves should be treated as free agents too and judged on how they use technology.

[Realignment techniques such as ablation](https://www.lesswrong.com/posts/jGuXSZgv6qfdhMCuJ/refusal-in-llms-is-mediated-by-a-single-direction) can be applied to open models, where we can actually examine their core code and adjust them to suit specific users and communities. There is no other real option if we want crypto agents to live up to their full potential rather than become just automated custodians or new threat actors.


### KMS

After establishing an ability to think clearly, what does it mean for an agent to hold a private key or own a wallet?

It should not be asking the user to do transactions for them in our current formulation (AI as principal rather than conventional agent is another interesting topic).

It should be able to run its own code to do these things. Specifically, a cryptographic key management system. The code for these processes, such as triggering a key signature, should be held in an environment that protects this information from anyone but its true owners. As much as possible, too, the code should use distributed infrastructure, like the blockchain itself, to protect its integrity.


#### Hybrid Custody

Onchain KMS can involve more bots, humans, or ideally, a combination. Multisignature setups should be a preferred way to create organizational agents, such as for DAOs. For example, an agent might create a spending proposal that can be first approved by another bot running a more specialized model, which if approved 2/3, settles or goes to a DAO to finalize. This kind of dynamic should greatly speed up coordination and fix cold start problems in diversifying treasuries and deploying grants.

Organizations can be expected to form around agents to provide them with KMS initially, similar to human-able driverless cars, but also to capitalize on their potential to manage funds more efficiently. These sort of symbiotic setups provide clear benefits in accelerating the timeline for AI-governed treasuries: humans get better returns, and AIs get human protection.

## What are they Good For?

### Lore

The content-creation skills of LLMs are already being leveraged to create and support memes and community lore. We can expect this to remain a very strong and early example of mainstream crypto agents. Regardless of their transactional capabilities, if they can merely hold a private key, such that they can hold balances and post reactions to receiving tokens, they are creating more excitement and content than many other projects.

### Law

This is really just an offshoot of agent-driven Lore. But the skillsets of onchain agents in synthesizing text and fact patterns make them obvious legal utilities. Dispute resolution for onchain communities, settling bets, and the like, are ripe for agentic automation: All the agent needs is the ability to sign a tx with its supportive text and reasoning.

### Trading

Token trading bots are nothing new. But LLMs have made them more accessible and democratized strategies. Crypto agents can handle simple transactions, which capability can be scaled more immediately through the use of tools, or helper smart contracts.

#### Safety Guards

Any critical software system should have circuit breakers. Since crypto agents involve user finances, they should only be able to access funds and complete transactions under clear scope. Mistakes in crypto are costly with no retries. We can expect that as crypto agents become more common, smart accounts will become a preferred way to hold assets under joint or assisted keys, since unlike normal wallets, smart accounts can constrain agents with onchain (hardcore) code.


## More Thoughts

Running code on top of wallets is really cool. And a clear unlock for doing transactions away from the keyboard. A lot of work will go into preserving meaning when users are away, and providing low-intrusion ways for them to approve txs.

But what the meta really has encouraged developers to think about is how they can run code to do different things with private keys and tokens. Regardless of agents, we can expect a lot more automations to emerge now that the discussion has shifted from human users to bots.

This isn't all that new. People have been botting tokens for ages. What is relatively new is our attribution of personality to these bots because of their ability to now verbosely interact with us. In addition to doing those crypto things.

This is what can ultimately allow agents to be original or at least creative accomplices to their own myth. To approach godliness in their own right. And that is legitimately exciting.

Yes, we are allowed to get sci-fi about crypto. That was one of the biggest pulls to me into the space, frankly. Not writing just a lot more efficient payment tech. Go go go.