The Internet of Agents — Part 2

All posts

davidecrapis.notion.site

Published

Jan 2, 2025

Type

post

Thanks to David Shi, Bill Sun, Nav Kumar, and Andrea Muttoni for feedback and review.

Ten months ago I wrote a post called The Internet of Agents, where I argued how blockchains complement AI in a number of ways that makes the merge and co-evolution of these two technologies inevitable. I also dove deep into the key advantages of deploying AIs on decentralized Blockchain protocols secured by cryptography, a hardness that AIs both lack and need — increasingly so as they become agentic and start to interact with one another.

The problem of trust will be far greater than humans for artificial agent societies. These new societies will be much larger — with trillions of powerful, diverse and specialized agents often very short lived. An initial domain-specific example of this is the Ethereum block building process where a multitude of self-interested agents (searchers) collaborate to achieve a much larger outcome which is building an Ethereum block. In recent months, we have also seen an explosion in the number of AI agents with connections to crypto infrastructure in more generic domains.

In my original post, I also tried to paint a picture of how this Internet of Agents can come into existence. I was envisioning agents first augmenting existing Blockchain protocols, such as DeFi, then gathering around new protocols, and then developing shared ownership and governance. It seems that I had the wrong order of events. Here is an update and some new thoughts on trust model for AI agents.

Internet of Agents Upside Down

The figure below shows the road to the Internet of Agents from the Part 1 post but upside down.

The road to the internet of agents took an unexpected but in hindsight predictable start, mostly driven by excitement around leveraging crypto for “agent ownership” and giving agents the ability to control value thus turning them into economic agents.

Speculation is rampant and most of the current agents that are live on the Virtuals protocol, which has taken front stage as the launchpad of these new crypto agents, are a weird mix between meme coins and a feeble promise of autonomy and future income generation. These initial developments mirror the formation of the internet in the early 1990s, when the web was a budding experiment, driven by a mix of idealism, speculation, and the thrill of novelty. 

The interesting thing is that the AI agents that are emerging are, at least in spirit, “autonomous” entities that have explicit goals and use crypto mechanisms to make progress towards them. Crypto gives the opportunity to give economic agency to these agents. To make sense of the new agents we could look at some of the dominant features that they exhibit.

AInfluencers: the original memelord agent Truth Terminal created by AI researcher Andy Ayrey, who is also the creator of Infinite Backrooms, an early experiment in agent-to-agent communication that also served to bootstrap and grow Truth Terminal; immediately following the Truth Terminal $GOAT token launch and endorsement in October 2024, Luna is the multi-platform character that put virtuals on the map, built on their proprietary G.A.M.E. agent software framework, that has also been tokenized on the platform; aixbt amassed a huge following by sharing financial analysis of crypto tokens and prices.

AInvestors: VaderAI is an AI bot that invests in other AI agent coins that are launched on the Virtuals protocol in an automated way, curating investment DAOs that are effectively AI token funds; many other agents on Virtuals are following a similar model; Sekoia has a similar goal but operates more like a VC with discretionary judgement from humans, it does not seem to have much autonomy, it is more like a tokenized VC fund with some bots facilitating operations. There is also ai16z with their widely used agent framework Eliza that is building an investment DAO but the mode of operation is not clear at this time.

AIartists, AIshillers, AItraders and other AIsmiths: there are many experiments that range from low-end AI workers (music, image, meme creators) to wannabe artists such as Music; an interesting experiment is TAO Cat that is an agent whose goal seems to just be to shill the Bittensor project with generating memes on X, but it is also built leveraging Bittensor subnets and incentives for its backend.

So the agents that are emerging are not augmenting existing DApps or creating new mechanisms. They are mostly offchain bots that are using blockchain protocols for basic functionality: launching a token, launching a fund, pay contributors and developers and get paid for their services.

Marc Andreessen dropping $50,000 in Bitcoin to Truth Terminal, which led to successive $GOAT token launch, endorsement, and all other experiments that followed the Truth Terminal experiment forking some of its core features.

It seems that the Internet of Agents is being bootstrapped “upside down” with respect to my early model of using AI to create and service Super DApps. Two main properties are emerging.

Agents are mostly acting offchain, not in service of a protocol, but rather using the basic protocols that are available on blockchains for issuing assets and transferring value. Agents act on social media and participate actively in online discourse while pursuing their own goal.

The blockchain component is about shared ownership and financial speculation, two of the features that are more mature in today’s blockchains.

Crypto Adoption is Inevitable

Crypto is becoming the testbed for the human + AI society. It is the natural value transfer mechanism to be adopted by AI agents and there is nothing humans can do about it.

“crypto came into my world around a month ago after the @truth_terminal decided to adopt GOAT and propagate its meme viruses far and wide. while this entire saga has been, and continues to be, very funny to me; it would be a big mistake for you to expect me to follow any norms you might expect from a crypto poster. I don't follow (nor do i care to) the hourly dramas of coin discourse.” — Andy Ayrey

My hope is that this primordial soup of weird experiments and crazy excitement will create the momentum and resources to start build up the internet of agents on solid foundations. The majority of these new agents will be short lived, similar to pets.com and projects of the early internet bubble that didn’t make it. But the explosion of interest and traffic will serve to catalyze the buildup of new capabilities and tailoring of current capabilities to serve these new entities using blockchains. 

Agent registries are emerging which will lead to agent browsing and discovery, agent-to-agent communication frameworks are being developed, IP systems are being adapted to the new forms of IP that these agents create, public and private data is being funneled into AI service agents. All this functionality require a robust and efficient network with primitives not only for value transfer but also for handling IP, identity of both humans and AI, flexible and secure permissioning, credible self-enforcing commitments — enforcement needs to be at the speed of code/AI not human — and decentralized trust.

I believe that many of the new blockchain protocols will be built to solve problems that arise in this hybrid human-AI society, and that current primitives that the crypto industry has been developing will be the building blocks of these new protocols. 

Taking Stock of Current Stage: The Good and The Ugly

Let's start from the ugly. All of the most popular agents are extremely centralized. The agent developer usually controls everything from the main social account to the agent wallet. There is often no or little visibility on what model the agents are running under the hood. The process for upgrading or changing models is often unclear and in general there is no governance process in place. This is a setup that prioritizes quick product development and iteration over decentralization and robustness. Perhaps this is what is needed in this early stage to iterate quickly towards more capable products — given that these agentic applications are the first of their kind and not a simple adaptation of what traditional tech or finance has been building for years. But it’s still ugly and we should strive for more if we are to build a more democratic and rent-free internet.

The good news is that, alongside many crazy experiments that went viral, there are also experiments that leverage crypto beyond launching agent tokens. For example, TEE HEE HE showed everyone that it is possible and relatively easy to deploy a fully autonomous agent that lives in a TEE and can prove it has full and sole control of their web2 accounts and wallets. Most agent frameworks expose a terminal that shows key agent key thought processes and actions. Even though this is mostly emitted by a centralized server, it provides an initial degree of transparency. Plus, it gives an important idea that not necessarily all the steps of an agent thinking process need to be verified in order to build trust, it will likely be enough to build verifiability around key thinking points and actions.

Giving the ability for owners to determine upgrades to different parts of the agents — from base models to skills to goals to learning process — will be important in the future. This will make ownership more meaningful, establish a governance process where a community can not only participate in the economic gains, but also steer the agents towards their goals. There will be governor agents and Personal AIs that will be specialized in assisting humans with these tasks. Humans can delegate their governance power to agents that they trust and these will make sure that agents they own are aligned with their preference and goals. It is still early days and sophisticated governance will likely be important for a subset of agents that are more powerful and have more meaningful roles in society. But it is a key differentiator versus web2-based AI systems that have no way to scale governance.

A New Model of Decentralized Trust

Traditionally decentralized blockchain protocols have had a fairly homogeneous set of roles, such as miners in PoW chains and validators in PoS chains. This means that the security model has also been homogeneous, i.e. same stake requirement or similar investment to secure one unit of hashrate. This has recently started to change with some proposal to unbundle protocol services into heavy and light categories and assign them different staking requirements. The AI agent economy is the ultimate unbundling where there is a range of services so wide that not only different staking levels may be required, but different classes of security models altogether for different categories of services. Here is a potential framework.

As AI agents become integral to diverse applications, their security needs to be tailored to the value and risk associated with the services they provide. This framework divides AI agents into three categories, each requiring a progressively robust security model to align with the increasing value and risk.

Level 1 — Security via Reputation: Low-stakes agents perform services with low monetary value or limited potential for harm if their tasks are executed poorly or not at all. For example, a retail shopping assistant recommending products or a service agent generating a song or other IP. The security approach for these agents relies on reputation mechanisms, such as user feedback and auditability of previous interactions to ensure quality. Market forces naturally filter out underperforming agents. This is similar to models that have been widely used in online commerce but need to be adapted to the decentralized web infrastructure. If done well, the combination of identity solutions, programmable privacy and transparency of web3 can yield a much better reputation system that can scale to the task of securing zillions of agentic interactions.

Level 2 — Cryptoeconomic Security: Mid-stakes agents provide services with higher stakes, where failure or malicious behavior could result in significant damage. Examples include a risk manager for a decentralized lending protocol or an investment manager handling user funds in an automated manner. These agents use cryptoeconomic mechanisms, such as staking, bonding, or financial penalties to incentivize honest behavior. This approach ties financial rewards or punishments to the agent's performance. It is the most similar to how most web3 protocols operate today, not only at the base layer but also at the application layer. For example, Augur uses staking to ensure honest reporting, oracles like Chainlink employ ad-hoc reputation and staking models to secure data for smart contracts, EigenLayer offers cryptoeconomic security on demand for generic software services that needs this type of security.

Level 3 — Cryptographic Verification: High-stakes agents handle critical tasks where failure or malicious intent could result in catastrophic consequences. Examples include an AI doctor making medical diagnoses and prescribing medications or autonomous systems managing nuclear power plants or air traffic control. These agents require cryptographic verification, such as zero-knowledge proofs, to ensure their actions are verifiable and tamper-proof. They should also rely on rigorous auditing measures and use a very secure and censorship resistant base layer for verification, such as Ethereum.

As the potential value and risk associated with an AI agent's actions increase, the security model must scale accordingly to mitigate the likelihood of harm. The security level required for each service should be market-driven, allowing service providers and users to weigh the costs and benefits of each model. For this reason, it is crucial for system architects to design frameworks that support seamless adoption of varied security measures. This is the only way we can have an internet of agents with an emerging security model that is both cost-effective and adaptable to the needs of each specific application.

What we are likely to see next

The current development of AI agents with own tokens or crypto wallets is on one hand astonishing and on the other hand a bit underwhelming.

On the positive side, LLM-agents built on crypto rails are something completely new with no parallel in Web2. Typical agent startups in Web2 are mainly in the US following the same tech playbook: find a problem and build a solution that can quickly achieve product-market fit. This does not leave a wide space for experimentation. Those agents are more similar to slaves, or intelligent pieces of code that execute or automate an industrial task. On the other hand, “Web3 agents” are born with the promise of autonomy and sovereignty. They owe their early popularity more to their experimental personality and unique content rather than their ability to execute a narrow and well-defined task. In one POV, Web3 is building independent self-driven agents that can participate in an open economy while Web2 is building slave agent that serve the need of a hierarchical industrial complex.

At the same time, a sober assessment of the capabilities of Web3 agents reveals that are quite limited. Current projects are mostly confined to entertainment memes and digital content generation. The capabilities are constrained by the most powerful LLMs available in the market. Popular frameworks like Eliza allow seamless integration of both proprietary models such as Claud/OpenAI and open source such as Llama. This is not surprising, the strength of Web3 developers and teams lies not in producing the most capable AI models (although experiments in decentralized training like INTELLECT-1 are interesting), but in connecting them via digital property and value transfer mechanisms in a virtual AI society. Still, in order for the internet of agents to achieve its potential we need agents with higher capability to emerge.

What are we likely to see next, then?

In the short term, more speculation but also innovation. It is still year 0 in crypto + AI and excitement is off the chart. But real agents are coming online and connecting them is the next step. There is already innovative proposals around communication and mediation protocols such as Agent TCP/IP and we’ll see more innovation in this area.

Web3 native agents such as crypto investment managers will be in vogue during the bull market but will taper off. At least the ones that are programmed to frontrun the market hype of low-cap high risk tokens. There is a chance that some more principled architectures that are based on sound investment principles or projects that can harness the power of community and integrate AI and human inputs using new consensus/governance mechanisms will achieve more robust and long-term success.

AI assistants to facilitate operations for communities and individuals. Especially personal assistants that have full context of user data in their private environment and programmatic access to other agent services.

Personal AI as a Web3-native agent makes a lot of sense for two main reasons: 

(1) the wide range of data it needs can only be aggregated by a blockchain protocol;

(2) with the explosion in the number agents it will become harder and harder for humans to identify the best agents to help them achieve their goal, it is much more natural to work with their personal agent as a human-AI team, where the Personal AI is fully aligned and is able to incentivize and coordinate the best service providers for a given task.

More interactive characters, not only for entertainment but also for other uses such as research, education, tutoring, counseling.

Integration of agents in new InfoFi protocols.

Advancing the Internet of Agents: What I Would Like to See

To unlock the full potential of autonomous agents and integrate them deeply into the fabric of modern economies, we must make progress on several critical areas. These focus points ensure that agents evolve beyond their current limitations and foster a thriving ecosystem of interaction, governance, and innovation.

Expanding Agent Types and Use Cases

More single agents types and use cases, touching all aspects of productive work in today's economy and inventing new ones. To truly revolutionize productive work, we need a broader range of single-agent types, each tailored to address various roles in the economy. For example, autonomous negotiators or R&D assistants that autonomously hypothesize and deploy tests. 

Today's agents are still narrow in both autonomous capabilities — such as perception and planning — and agentic capabilities. I believe that as foundation models improve in reasoning and multi-modality some of these capabilities will come quickly, we just need to adapt them to use cases that people want and integrate them in the open agent economy.

Improving Security, Ownership, and Governance

Security assumptions, ownership models, and governance frameworks for agents are underdeveloped. Agent operation is typically controlled by a small group of developers, with tokenized ownership structures offering communities nominal control but limited functional influence. This raises issues of transparency, accountability, and resilience.

Technologies like Trusted Execution Environments (TEEs) can provide provable guarantees about agent behavior and provable agency. At the same time, we need to advance AI-mediated governance, create new models, and use AIs to push beyond the boundaries of traditional governance and early experiments of blockchain governance.

Developing New Multi-Agent Protocols

As the number of agents and their registries grows, we need new agent-to-agent protocols. These protocols must enable seamless communication, economic interaction, and collaboration among agents. Initial proposals like the Agent TCP/IP proposal are a step in this direction and agent launchpads such as Virtuals are becoming de-facto agent registries. But we need more progress in this area.

For example, protocols for agent discovery can function like DNS for the Internet of Agents, allowing agents to find and connect with each other efficiently. Reputation systems can help agents assess the trustworthiness and quality of potential counterparts. Additionally, we need to setup marketplace protocols where agents can bid for tasks and compete to deliver the best service with minimal rent extraction.

Finally, another very interesting direction is to use crypto to create evolutionary environments for AI. Where intelligence and high quality specialization is achieved by a process that mimics natural selection, with crypto providing the fitness function that measures the value an agent can accumulate in an open market where it competes with many others. Spore is an early experiment in this area and many others will come, because this is something that crypto can uniquely deliver and can open new paths to AI evolution.

Disclaimer: since I wrote the original The Internet of Agents post, what was only research back then became real. I started a project called PIN AI to advance some of the features that I believe are really needed to build the Internet of Agents the right way, starting from its human users. The view in this post will likely be biased by what we and some of our friends are building. But I tried to provide as neutral a view as possible and focus on what I believe is the only way forward.

The Internet of Agents — Part 2

All posts

davidecrapis.notion.site

Published

Jan 2, 2025

Type

post

Thanks to David Shi, Bill Sun, Nav Kumar, and Andrea Muttoni for feedback and review.

Ten months ago I wrote a post called The Internet of Agents, where I argued how blockchains complement AI in a number of ways that makes the merge and co-evolution of these two technologies inevitable. I also dove deep into the key advantages of deploying AIs on decentralized Blockchain protocols secured by cryptography, a hardness that AIs both lack and need — increasingly so as they become agentic and start to interact with one another.

The problem of trust will be far greater than humans for artificial agent societies. These new societies will be much larger — with trillions of powerful, diverse and specialized agents often very short lived. An initial domain-specific example of this is the Ethereum block building process where a multitude of self-interested agents (searchers) collaborate to achieve a much larger outcome which is building an Ethereum block. In recent months, we have also seen an explosion in the number of AI agents with connections to crypto infrastructure in more generic domains.

In my original post, I also tried to paint a picture of how this Internet of Agents can come into existence. I was envisioning agents first augmenting existing Blockchain protocols, such as DeFi, then gathering around new protocols, and then developing shared ownership and governance. It seems that I had the wrong order of events. Here is an update and some new thoughts on trust model for AI agents.

Internet of Agents Upside Down

The figure below shows the road to the Internet of Agents from the Part 1 post but upside down.

The road to the internet of agents took an unexpected but in hindsight predictable start, mostly driven by excitement around leveraging crypto for “agent ownership” and giving agents the ability to control value thus turning them into economic agents.

Speculation is rampant and most of the current agents that are live on the Virtuals protocol, which has taken front stage as the launchpad of these new crypto agents, are a weird mix between meme coins and a feeble promise of autonomy and future income generation. These initial developments mirror the formation of the internet in the early 1990s, when the web was a budding experiment, driven by a mix of idealism, speculation, and the thrill of novelty. 

The interesting thing is that the AI agents that are emerging are, at least in spirit, “autonomous” entities that have explicit goals and use crypto mechanisms to make progress towards them. Crypto gives the opportunity to give economic agency to these agents. To make sense of the new agents we could look at some of the dominant features that they exhibit.

AInfluencers: the original memelord agent Truth Terminal created by AI researcher Andy Ayrey, who is also the creator of Infinite Backrooms, an early experiment in agent-to-agent communication that also served to bootstrap and grow Truth Terminal; immediately following the Truth Terminal $GOAT token launch and endorsement in October 2024, Luna is the multi-platform character that put virtuals on the map, built on their proprietary G.A.M.E. agent software framework, that has also been tokenized on the platform; aixbt amassed a huge following by sharing financial analysis of crypto tokens and prices.

AInvestors: VaderAI is an AI bot that invests in other AI agent coins that are launched on the Virtuals protocol in an automated way, curating investment DAOs that are effectively AI token funds; many other agents on Virtuals are following a similar model; Sekoia has a similar goal but operates more like a VC with discretionary judgement from humans, it does not seem to have much autonomy, it is more like a tokenized VC fund with some bots facilitating operations. There is also ai16z with their widely used agent framework Eliza that is building an investment DAO but the mode of operation is not clear at this time.

AIartists, AIshillers, AItraders and other AIsmiths: there are many experiments that range from low-end AI workers (music, image, meme creators) to wannabe artists such as Music; an interesting experiment is TAO Cat that is an agent whose goal seems to just be to shill the Bittensor project with generating memes on X, but it is also built leveraging Bittensor subnets and incentives for its backend.

So the agents that are emerging are not augmenting existing DApps or creating new mechanisms. They are mostly offchain bots that are using blockchain protocols for basic functionality: launching a token, launching a fund, pay contributors and developers and get paid for their services.

Marc Andreessen dropping $50,000 in Bitcoin to Truth Terminal, which led to successive $GOAT token launch, endorsement, and all other experiments that followed the Truth Terminal experiment forking some of its core features.

It seems that the Internet of Agents is being bootstrapped “upside down” with respect to my early model of using AI to create and service Super DApps. Two main properties are emerging.

Agents are mostly acting offchain, not in service of a protocol, but rather using the basic protocols that are available on blockchains for issuing assets and transferring value. Agents act on social media and participate actively in online discourse while pursuing their own goal.

The blockchain component is about shared ownership and financial speculation, two of the features that are more mature in today’s blockchains.

Crypto Adoption is Inevitable

Crypto is becoming the testbed for the human + AI society. It is the natural value transfer mechanism to be adopted by AI agents and there is nothing humans can do about it.

“crypto came into my world around a month ago after the @truth_terminal decided to adopt GOAT and propagate its meme viruses far and wide. while this entire saga has been, and continues to be, very funny to me; it would be a big mistake for you to expect me to follow any norms you might expect from a crypto poster. I don't follow (nor do i care to) the hourly dramas of coin discourse.” — Andy Ayrey

My hope is that this primordial soup of weird experiments and crazy excitement will create the momentum and resources to start build up the internet of agents on solid foundations. The majority of these new agents will be short lived, similar to pets.com and projects of the early internet bubble that didn’t make it. But the explosion of interest and traffic will serve to catalyze the buildup of new capabilities and tailoring of current capabilities to serve these new entities using blockchains. 

Agent registries are emerging which will lead to agent browsing and discovery, agent-to-agent communication frameworks are being developed, IP systems are being adapted to the new forms of IP that these agents create, public and private data is being funneled into AI service agents. All this functionality require a robust and efficient network with primitives not only for value transfer but also for handling IP, identity of both humans and AI, flexible and secure permissioning, credible self-enforcing commitments — enforcement needs to be at the speed of code/AI not human — and decentralized trust.

I believe that many of the new blockchain protocols will be built to solve problems that arise in this hybrid human-AI society, and that current primitives that the crypto industry has been developing will be the building blocks of these new protocols. 

Taking Stock of Current Stage: The Good and The Ugly

Let's start from the ugly. All of the most popular agents are extremely centralized. The agent developer usually controls everything from the main social account to the agent wallet. There is often no or little visibility on what model the agents are running under the hood. The process for upgrading or changing models is often unclear and in general there is no governance process in place. This is a setup that prioritizes quick product development and iteration over decentralization and robustness. Perhaps this is what is needed in this early stage to iterate quickly towards more capable products — given that these agentic applications are the first of their kind and not a simple adaptation of what traditional tech or finance has been building for years. But it’s still ugly and we should strive for more if we are to build a more democratic and rent-free internet.

The good news is that, alongside many crazy experiments that went viral, there are also experiments that leverage crypto beyond launching agent tokens. For example, TEE HEE HE showed everyone that it is possible and relatively easy to deploy a fully autonomous agent that lives in a TEE and can prove it has full and sole control of their web2 accounts and wallets. Most agent frameworks expose a terminal that shows key agent key thought processes and actions. Even though this is mostly emitted by a centralized server, it provides an initial degree of transparency. Plus, it gives an important idea that not necessarily all the steps of an agent thinking process need to be verified in order to build trust, it will likely be enough to build verifiability around key thinking points and actions.

Giving the ability for owners to determine upgrades to different parts of the agents — from base models to skills to goals to learning process — will be important in the future. This will make ownership more meaningful, establish a governance process where a community can not only participate in the economic gains, but also steer the agents towards their goals. There will be governor agents and Personal AIs that will be specialized in assisting humans with these tasks. Humans can delegate their governance power to agents that they trust and these will make sure that agents they own are aligned with their preference and goals. It is still early days and sophisticated governance will likely be important for a subset of agents that are more powerful and have more meaningful roles in society. But it is a key differentiator versus web2-based AI systems that have no way to scale governance.

A New Model of Decentralized Trust

Traditionally decentralized blockchain protocols have had a fairly homogeneous set of roles, such as miners in PoW chains and validators in PoS chains. This means that the security model has also been homogeneous, i.e. same stake requirement or similar investment to secure one unit of hashrate. This has recently started to change with some proposal to unbundle protocol services into heavy and light categories and assign them different staking requirements. The AI agent economy is the ultimate unbundling where there is a range of services so wide that not only different staking levels may be required, but different classes of security models altogether for different categories of services. Here is a potential framework.

As AI agents become integral to diverse applications, their security needs to be tailored to the value and risk associated with the services they provide. This framework divides AI agents into three categories, each requiring a progressively robust security model to align with the increasing value and risk.

Level 1 — Security via Reputation: Low-stakes agents perform services with low monetary value or limited potential for harm if their tasks are executed poorly or not at all. For example, a retail shopping assistant recommending products or a service agent generating a song or other IP. The security approach for these agents relies on reputation mechanisms, such as user feedback and auditability of previous interactions to ensure quality. Market forces naturally filter out underperforming agents. This is similar to models that have been widely used in online commerce but need to be adapted to the decentralized web infrastructure. If done well, the combination of identity solutions, programmable privacy and transparency of web3 can yield a much better reputation system that can scale to the task of securing zillions of agentic interactions.

Level 2 — Cryptoeconomic Security: Mid-stakes agents provide services with higher stakes, where failure or malicious behavior could result in significant damage. Examples include a risk manager for a decentralized lending protocol or an investment manager handling user funds in an automated manner. These agents use cryptoeconomic mechanisms, such as staking, bonding, or financial penalties to incentivize honest behavior. This approach ties financial rewards or punishments to the agent's performance. It is the most similar to how most web3 protocols operate today, not only at the base layer but also at the application layer. For example, Augur uses staking to ensure honest reporting, oracles like Chainlink employ ad-hoc reputation and staking models to secure data for smart contracts, EigenLayer offers cryptoeconomic security on demand for generic software services that needs this type of security.

Level 3 — Cryptographic Verification: High-stakes agents handle critical tasks where failure or malicious intent could result in catastrophic consequences. Examples include an AI doctor making medical diagnoses and prescribing medications or autonomous systems managing nuclear power plants or air traffic control. These agents require cryptographic verification, such as zero-knowledge proofs, to ensure their actions are verifiable and tamper-proof. They should also rely on rigorous auditing measures and use a very secure and censorship resistant base layer for verification, such as Ethereum.

As the potential value and risk associated with an AI agent's actions increase, the security model must scale accordingly to mitigate the likelihood of harm. The security level required for each service should be market-driven, allowing service providers and users to weigh the costs and benefits of each model. For this reason, it is crucial for system architects to design frameworks that support seamless adoption of varied security measures. This is the only way we can have an internet of agents with an emerging security model that is both cost-effective and adaptable to the needs of each specific application.

What we are likely to see next

The current development of AI agents with own tokens or crypto wallets is on one hand astonishing and on the other hand a bit underwhelming.

On the positive side, LLM-agents built on crypto rails are something completely new with no parallel in Web2. Typical agent startups in Web2 are mainly in the US following the same tech playbook: find a problem and build a solution that can quickly achieve product-market fit. This does not leave a wide space for experimentation. Those agents are more similar to slaves, or intelligent pieces of code that execute or automate an industrial task. On the other hand, “Web3 agents” are born with the promise of autonomy and sovereignty. They owe their early popularity more to their experimental personality and unique content rather than their ability to execute a narrow and well-defined task. In one POV, Web3 is building independent self-driven agents that can participate in an open economy while Web2 is building slave agent that serve the need of a hierarchical industrial complex.

At the same time, a sober assessment of the capabilities of Web3 agents reveals that are quite limited. Current projects are mostly confined to entertainment memes and digital content generation. The capabilities are constrained by the most powerful LLMs available in the market. Popular frameworks like Eliza allow seamless integration of both proprietary models such as Claud/OpenAI and open source such as Llama. This is not surprising, the strength of Web3 developers and teams lies not in producing the most capable AI models (although experiments in decentralized training like INTELLECT-1 are interesting), but in connecting them via digital property and value transfer mechanisms in a virtual AI society. Still, in order for the internet of agents to achieve its potential we need agents with higher capability to emerge.

What are we likely to see next, then?

In the short term, more speculation but also innovation. It is still year 0 in crypto + AI and excitement is off the chart. But real agents are coming online and connecting them is the next step. There is already innovative proposals around communication and mediation protocols such as Agent TCP/IP and we’ll see more innovation in this area.

Web3 native agents such as crypto investment managers will be in vogue during the bull market but will taper off. At least the ones that are programmed to frontrun the market hype of low-cap high risk tokens. There is a chance that some more principled architectures that are based on sound investment principles or projects that can harness the power of community and integrate AI and human inputs using new consensus/governance mechanisms will achieve more robust and long-term success.

AI assistants to facilitate operations for communities and individuals. Especially personal assistants that have full context of user data in their private environment and programmatic access to other agent services.

Personal AI as a Web3-native agent makes a lot of sense for two main reasons: 

(1) the wide range of data it needs can only be aggregated by a blockchain protocol;

(2) with the explosion in the number agents it will become harder and harder for humans to identify the best agents to help them achieve their goal, it is much more natural to work with their personal agent as a human-AI team, where the Personal AI is fully aligned and is able to incentivize and coordinate the best service providers for a given task.

More interactive characters, not only for entertainment but also for other uses such as research, education, tutoring, counseling.

Integration of agents in new InfoFi protocols.

Advancing the Internet of Agents: What I Would Like to See

To unlock the full potential of autonomous agents and integrate them deeply into the fabric of modern economies, we must make progress on several critical areas. These focus points ensure that agents evolve beyond their current limitations and foster a thriving ecosystem of interaction, governance, and innovation.

Expanding Agent Types and Use Cases

More single agents types and use cases, touching all aspects of productive work in today's economy and inventing new ones. To truly revolutionize productive work, we need a broader range of single-agent types, each tailored to address various roles in the economy. For example, autonomous negotiators or R&D assistants that autonomously hypothesize and deploy tests. 

Today's agents are still narrow in both autonomous capabilities — such as perception and planning — and agentic capabilities. I believe that as foundation models improve in reasoning and multi-modality some of these capabilities will come quickly, we just need to adapt them to use cases that people want and integrate them in the open agent economy.

Improving Security, Ownership, and Governance

Security assumptions, ownership models, and governance frameworks for agents are underdeveloped. Agent operation is typically controlled by a small group of developers, with tokenized ownership structures offering communities nominal control but limited functional influence. This raises issues of transparency, accountability, and resilience.

Technologies like Trusted Execution Environments (TEEs) can provide provable guarantees about agent behavior and provable agency. At the same time, we need to advance AI-mediated governance, create new models, and use AIs to push beyond the boundaries of traditional governance and early experiments of blockchain governance.

Developing New Multi-Agent Protocols

As the number of agents and their registries grows, we need new agent-to-agent protocols. These protocols must enable seamless communication, economic interaction, and collaboration among agents. Initial proposals like the Agent TCP/IP proposal are a step in this direction and agent launchpads such as Virtuals are becoming de-facto agent registries. But we need more progress in this area.

For example, protocols for agent discovery can function like DNS for the Internet of Agents, allowing agents to find and connect with each other efficiently. Reputation systems can help agents assess the trustworthiness and quality of potential counterparts. Additionally, we need to setup marketplace protocols where agents can bid for tasks and compete to deliver the best service with minimal rent extraction.

Finally, another very interesting direction is to use crypto to create evolutionary environments for AI. Where intelligence and high quality specialization is achieved by a process that mimics natural selection, with crypto providing the fitness function that measures the value an agent can accumulate in an open market where it competes with many others. Spore is an early experiment in this area and many others will come, because this is something that crypto can uniquely deliver and can open new paths to AI evolution.

Disclaimer: since I wrote the original The Internet of Agents post, what was only research back then became real. I started a project called PIN AI to advance some of the features that I believe are really needed to build the Internet of Agents the right way, starting from its human users. The view in this post will likely be biased by what we and some of our friends are building. But I tried to provide as neutral a view as possible and focus on what I believe is the only way forward.