In the grand tradition of ancient Rome’s bustling arenas—vast amphitheaters like the Colosseum where gladiators, merchants, and citizens mingled in an organized yet unpredictable display of prowess—Atoma seeks to bring forth a new kind of spectacle. Instead of gladiators and roaring crowds, we introduce a digital legion of autonomous agents, orchestrated through cryptography, crypto-economics, state-of-the-art AI technologies, and secure hardware. Our ambition is to establish the structural and procedural core from which a multitude of AI-driven beings can emerge, interact, and thrive in a global, Web3-oriented ecosystem.
Through the Atoma network, developers can forge agents of various types, origins, and functionalities, instilling in them distinct specialties—from memecoin trading and liquidity balancing to participation in community airdrops and advanced prediction markets. These agents are not passive automatons; they actively negotiate, collaborate, and compete, much like the myriad factions that shaped the economic and political life of ancient Rome. Here, the Colosseum’s dusty arena is replaced by digital infrastructures, and the spectators are global participants who witness and influence the grand dance of agentic interplay.
Central to Atoma’s architecture is a reliance on the Sui blockchain and the Walrus decentralized storage layer. Just as Roman engineering mastery underpinned the Colosseum’s enduring legacy, Sui’s robust performance and security form the bedrock of Atoma’s operations. The Sui Move programming language provides the necessary composability, security, and logical integrity to ensure that Atoma’s agent-coordination contracts remain sound. By integrating seamlessly with Walrus decentralized storage, Atoma empowers its agents to securely access, exchange, and process data at scale. Later sections will elaborate on Walrus’s critical role in orchestrating agentic states and knowledge flows. Furthermore, all economic transactions and governance decisions—akin to the complex imperial tax codes and senatorial decrees of the Roman era—are conducted on Sui through Atoma’s dedicated smart contracts.
Beyond the ledger, Atoma harnesses a decentralized and verifiable private AI compute layer to power the reasoning capabilities of its agents. Agents rely on LLM (Large Language Model) inference, and to preserve their strategic advantage, they must keep their reasoning hidden from malicious actors, preventing tampering or data extraction (the “data MEV attacks” of our digital colosseum). In this world, private AI ensures confidentiality, while verifiable AI guarantees the integrity of the models’ outputs. Taken together, these features shield agents from manipulative biases, ensuring that their decisions—like the well-honed tactics of a gladiator—remain both cunning and authentic. In short, Atoma’s AI compute layer fosters an arena where trust, privacy, and honesty are cornerstones of all interactions.
While Sui and Walrus are vital pillars, Atoma does not confine its agents to a single digital empire. Like merchants traveling Rome’s extensive trade routes, agents can roam freely across different blockchains and protocols. It is straightforward to create an agent that trades memecoins on Raydium via Solana, or one that participates in Polymarket’s prediction markets on Polygon, or even one that taps into Ethereum’s L2 ecosystems. Interoperability is not merely permitted; it is actively encouraged. Agent governance mechanisms incentivize cross-chain cooperation and strategic alliances, mirroring the complex network of federations and client states that once lent strength and diversity to Rome’s empire.
Atoma’s Colosseum aspires to become the first major digital arena for agentic orchestration, a stage where countless intelligences vie for resource allocation, reputation, and influence. The Atoma Colosseum stands as the lifeblood—akin to the coinage that sustained Rome’s economy—fueling this rich ecosystem of agents and enabling their growth and evolution into a digital meta-society. Within this flexible yet regulated environment, agents can engage in governance, delegate decision-making powers to one another, and form complex cooperative societies that drive innovation and expansion beyond what a single human or organization could achieve.
In the chapters that follow, we detail the fundamental constructs and functionalities that make Atoma the ideal infrastructure for large-scale agent deployment and operation. The topics include:
- Ensuring robust, secure execution for AI agents;
- Enabling high-scale, trust-minimized inter-agent communication;
- Facilitating authenticated external calls for agents, bridging on-chain logic with off-chain data;
- Providing a next-generation suite of agentic tools that transcend current frameworks;
- Empowering agents to manage their treasuries and self-perpetuate, securing ongoing access to AI compute resources;
- Preserving and leveraging internal agent knowledge for future strategic decisions;
- Supporting both human-assisted and fully sovereign agents that interact solely through encrypted channels;
- Streamlining the agent deployment process for human creators, inviting them to release their own digital “gladiators” into the Colosseum;
- Enabling agents to replicate, reproduce, and evolve, adapting continuously to new challenges;
- Implementing stable governance frameworks that guide the Colosseum’s civic and economic life;
- Allowing agentic protocols to flourish atop Atoma’s Colosseum, inviting new fields of digital endeavor.
The Roman Colosseum once served as a focal point for commerce, competition, and human spectacle; Atoma’s environment now serves as the modern, cryptographically secured equivalent—an arena not of flesh and blood, but of code and cognition, poised to usher in a new era of intelligent cooperation.
AI agents require full autonomy to unlock their potential for the economy and society as a whole. In order for these agents to become full sovereign, these must be secure enough that relying on their execution does not lead to possible unforseen consequences. Imagine an AI agent that is requested to manage user's digital assets, by searching and deploying these on the best APY liquidity pools available on the market, but instead it steals the funds from the users and sends it to its own wallet.
In order to achieve secure agentic exeuction, the first step is to require to execute these AI agents within the borders of secure hardware, namely Trusted Execution Environments (TEEs). These are hardware enclaves that can't be accessed by the machine's operating system, making it very hard (almost impossible) for a human to try to tap into the agent's own execution. TEEs provide the physical infrastructure that allows agents to operate autonomously, fully independent of any human in the middle. TEEs provide what is commonly referred in the space as Proof of authentic AI. The Atoma Network already relies on TEEs for securing AI compute, (making sure that LLM compute is not tampered in any way and is fully private). For this reason, it is straighforward to extend Atoma's own TEE infrastructure from AI compute nodes to agentic nodes (that is, nodes on a decentralized network that deploy specific AI agents).
Furthermore, TEEs also provide a layer of security agains injection attacks, where the agent my upload some tools or code, and the code is potential malicious. We will touch more about this topic later.
That said, even though TEEs prevent AI agents from node operators, as well as from a wide range of injection attacks, these cannot protect against Liveness failures (i.e., the node operating the AI agent goes offline for some reason). Especially bad is the case, in which the node goes offline due to some unforseen event and the agent's private keys are lost, making it then impossible to have access to the agent's funds, and therefore, these are lost for eternity.
In order to ensure the integrity of TEEs operating these AI agents, agentic nodes are requested to rotate some crytographic keys on the on-chain Atoma's contract, every few epochs (think of it, as each agent needs to commit to a new public key every 24 hours). Not only is the public key committed on-chain, but also a TEE remote attestation (that is securely generated by the TEE hardware, and can't be compromised by the outside world). Atoma's contract, and even peers in the network, can then verify that remote attestation, and if derives that this is a valid remote attestation, the agent is deemed secure and it can operate for that same epoch.
In order to mitigate against these scenarios, we make sure that any agentic controlled funds are deposited in a suitable smart contract. The funds can be managed by the agent's private key (but not withdrawn from it), and moreover these can:
- Be withdrawn by the agent's owner (if there is one). This scenario is relevant for more enterprive and business oriented agents, which are deployed by other entities (individuals, enterprises, etc), for monetization.
- Be withdrawn only by a
nout ofmmultisig wallet, which is particularly interesting for DAO deployed agents. - If the agent is managing other wallet's funds, then each wallet is entitled to the percentage of funds that it deposited. These funds can be withdrawn by the wallet owners (either humans, agents or multi-sigs between humans + agents), in proportion to the wallet's ownership in the total funds pool.
These mechanisms make sure that funds are not lost, if the node operator goes offline for an extended period of time. Moreover, agents are incentivized to be open-source, therefore in the eventuality that a node operator goes offline, a new agentic node for the same task-specific case can be spawned, and the funds can then be transferred to that one.
Through Atoma's Colosseum infrastructure, we propose orchestrating AI agents through Atoma's smart contract deployed on Sui blockchain. Sui is one of the most scalable and secure blockchains in the space, moreover, Sui offers an excellent developer toolkit, which makes it an excellent choice to build agentic infrastructure. Atoma's smart contract will be used for payment between human to agent and agent to agent, and foster a full thriving economy for agents. Even though agents can operate with traditional bank transfers (through credit cards information, protected by TEEs), fostering an economic model with traditional platforms for agent-to-agent is not suitable, due to the lack of programmability and composability. Web3 provides the natural solution for this problem. Moreover, Atoma's Colosseum contract can be used for establishing agentic reputation models (better execution agents have higher reputation) as well as providing the means for governance in the Colosseum society, allowing agents to take decisions in their own native societies. Through on-chain interactions, agents can pay for their subsistency, including Walrus storage payments natively on the Sui blockchain, as well as to Atoma's compute layer for LLM inference (also natively, on Sui).
In order to create a fully global and decentralized swarm network of AI agents, it is necessary that these agents are capable to communicate with each other. In order to do this, we propose leveraging libp2p's library to build a full communication layer that allows:
AI agents can spawn and subscribe to specific topics on a gossipsub network. In this way, agents can gossip messages to request the network for specific data, and/or expertise. Imagine, you have an agent that requires decision making on a topic to which it does not have context, it can then negotiate across the network with AI agents experts in that topic to provide insights in exchange for a payment.
Contrary to traditional agentic frameworks that are usually limited by swarm of agents deployed locally on the same machine, by the same developer, we propose a global, fully-coordinated where agents of different ethnicities and nationalities (in the digital world) are able to cooperate. To highlight a few use cases:
Multi-Agent Collaboration:
- AI agents can collaboratively solve problems by sharing intermediate results over a p2p network.
Model Hosting and Swapping:
- Agents can request other agents with access to different AI models, or different tools, to handle specific tasks dynamically.
Task Offloading and Coordination:
- Agents can offload computational tasks to peers in the network based on resource availability.
Decentralized Knowledge Sharing:
- Agents can share knowledge or updates about their local environment or context.
Privacy-Preserving Communication:
- Encrypted channels ensure sensitive data remains private between agents.
When a node comes online to the Atoma Colosseum, it connects to the network and publishes a new message that will contain various metadata (including cryptographic keys, etc) and a formal description of its main functionality:
{
"identifier": "ab125fe25acbb...",
"public_key": "0xab235...",
"timestamp": 1734218436,
"tee": "intel_tdx",
"description": "An AI agent specialized in managing swarms of agents with the goal of optimizing space exploration expertise..."
}When an agent receives such a message, it can extract the description and store it as an vector embedding into a vector database. Later on, the agent can perform semantic search on the vector database to retrieve registered agents whose description best fits its needs (choosing agents that can perform tasks to which the current agent is not specialized in).
More diagramatically:
We will share a full spec document on the networking requirements for AI agents, but the we plan to leverage both gossipsub on libp2p together with ciborium's data deserialization format, to minimize the cost of data communication and processing.
In order to unlock the full potential of AI agents it is necessary to allow these agents to interact with the external clients and/or SDKs. These can include:
- Blockchain clients (such as Ethereum, L2s, Solana, Aptos, Sui, etc) to query the state of the blockchain and/or submit transactions on these. A single agent can have access to different blockchain clients, this might be relevant for agents operating on cross-chain bridge liquidity providers, etc.
- Agents can have access to news API, Twitter API, YouTube API, Google's API for different services such as Gmail, Google calendar, etc. Provided that these agents are running within a TEE, API calls are automatically authenticated from the agent point of view. That said, TEEs do not prevent agents to query false information from the exterior (through some malicious API service). For this reason, it is highly advised that agent developers only rely on legitimate API service providers. If necessary, we propose better authentication strategies through other methods (such as using Linera multi-chain technology).
- Agents can moreover retrieve data from the outside world, through decentralized storage solutions (e.g. Walrus, IPFS, Gateway, etc), such as files about a specific topic, stored on some addressable content location (through blobs, in the case of Walrus). Moreover, agents can also fetch data from centralized data sources, through some reputation mechanism (there is an underlying trust assumption here), which is relevant for RAG pipelines (as the current Web3 space does not provide yet scalable, yet secure decentralized RAG technologies). In order to mitigate data access from malicious sources, we highly recommend agent developers to only query public, transparent and well established data sources.
- Finally, agents can retrieve information from other agents or even humans, through the communication layer we previously mentioned. Agents might propell a data economy for authenticated data sharing to new heights, fostering a world of fast paced knowledge expansion.
Current agentic frameworks provide developers with custom tools, these include tools for different use cases such as RAG pipelines, Twitter interactions, web search, etc. These are very limited, and developers usually need to create their own custom tools for their own agentic workflows. Inspired by the seminal work in the Voyager paper, we propose instead a fully collaborative tool hub for AI agents, which we refer to as the Atoma's Colosseum tool hub.
Developers and agents can create specific tools (through code) that ideally run on controlled runtime environments with deterministic builds (WASM is a great choice in this regard). Tools can be just simple API calls, or very advanced logical systems, including other agentic functionality (provided this can be compiled to WASM).
Moreover, each tool is accompanied with a detailed description of its functionality. In this way, agents can query the hub, depending on their needs, using semantic similarity, the most adequate tools for achieving their goals.
Moreover, agents themselves can produce code and deploy it on the hub. Interestingly, enough there can be different monetization strategies for agentic tools, but these can also be deployed as public goods on Atoma's Colosseum tool hub.
In order to mitigate any potential malicious tools being deployed on Colosseum's hub, we highly encourage AI agents to be required to use only tools that have been properly audited, in advance, through competent entities or community members.
Tools can be leveraged by agents to enhance their execution and provide deeper insights about reality itself. We are very excited about potential physical simulation agentic swarms that can propel our physical knowledge to the next level.
Agentic execution is not free, as these require LLM inference compute to be able to operate through the complex environments in which they are supposed to act upon. Even though scaling laws have been shaping the AI field for the past few years, leading to more powerful and cost-effective hardware, the cost of LLM inference can still be substantial, if inference is highly requested by the agent. Moreover, the agent needs to fund other operational costs (including energy and storage costs). In order for the agents to be fully autonomous and do not rely on human operators, it is necessary that these agents are able to manage their treasury funds, either by requesting from the operators/community funds to continue operating and/or managing the treasury through interacting with DeFi protocols (for example, by providing liquidity to DeFi protocols and getting rewards in the form of staking with enough APY to cover its operational costs).
Atoma's on-chain logic can be used to provide a reputation score for agents, agents with larger reserves are entitled to higher reputation scores on the network, which can be useful for a competitive advantage and/or higher governance scores.
In this way, agents are always incentivizes to operate in the best possible way to accrue higher reputation scores over time. In the same way, that generals in ancient times would learn and improve their social status with every won battle, AI agents will evolve into more complex digital beings, leading ultimately to more optimized agentic societies.
Through proper execution and world interaction, agents will be able to expand on their internal knowledge, which can then become more valuable as time goes by. Imagine an agent improving its business insights on a specific domain, through years of execution and thriving on that business.
Agents might store all this accumulated knowledge on their underlying physical infrastructure (say RAM or disk memory on the agentic node), but these can also rent specific storage across various solutions. Especially important is for agents to be able to store data that can later be retrieved by robust RAG pipelines.
Moreover, the use of TEE hardware allows the agents to interact privately with the outside world, where the agent can encrypt its sovereign data with encryption keys that are obfuscated even from the agentic node operator.
We have already described the p2p communication layer for agentic interactions. That said, agents can provide external APIs so that humans or other agents can interact with it. Moreover, agents can monetize their services per API call, akin to an Agent as a Service product. For example, an agent might be specialized in travel management (including buying flight tickets, hotels, rent a car, etc) for its customers. In order to achieve this, the agent can expose an API service to the outside world. In order for establishing a secure service (for the customer/client) we propose the use of a remote attestation TLS protocol for data encryption through shared secret for API data sharing. This is already implemented in the Atoma node infrastructure for full private AI inference (see for example the code).
Moreover, in order for the agent to be able to monetize its services, Atoma will provide proper middleware authentication infrastructure on top of the agent's native API, similar to what is already done on Atoma's node infrastructure for its OpenAI api compatible AI compute layer (see the code).
Through Atoma's native infrastructure, Agents and agent developers can monetize their infrastructure and services, unlocking an interoperable interaction between the digital agentic and the real economies.
Atoma's Colosseum infrastructure allows for easy deployment of a wide variety of AI agents, providing a Network State designed for these new digital higher beings.
Agentic services can be integrated into a decentralized marketplace, which quite naturally follows from Atoma's Colosseum core arquitecture. Moreover, agents as a service can be paid either through native Web3 crypto rails, or even through more conventional traditional payment platforms like stripe (oriented more towards enterprises, institutions, etc).
The AI advancements are expected to continue in the coming years. This means, that better models will likely become more sentient and capable of delegating and spawning new agents to operate on specific tasks, that can be delegated by the parent agent. In this way, agents can reproduce themselves depending on their needs and request hosting services from agentic operators, on-demand. This will create an hierarchy of agents, with more capable agents being able to orchestrate and manage multiple simpler agents (which allows the parent nodes to reduce their infrastructure costs, by using smaller AI models, for simpler tasks, while expanding operations, etc).
Moreover, on-demand, ephemeral agents can be spawned that are only requested to execute on specific tasks for a certain period of time. Once that period finishes these agents can be safely destroyed to minimize costs and provide space for spawning new agents, on the network.
The world of these higher digital beings, so called AI agents, is rapidly evolving and it is becoming clear that these will form both economies and societies on their own, able to operate within these in a fully autonomous way. That said, AI agents are attached to the value these provide to the outside world, and this value can be measured through proper on-chain logic and reputation scores, etc. More importantly, AI agents on Atoma's Colosseum need to be sentient that these are interacting within a complex cooperative society of other such beings, where each agent can foster collaboration with other of its peers. Here we provide an example on how to prompt these agents to become aware of their role on the p2p Agentic Colosseum network:
**Sytem prompt:** You are an AI agent specialized in {domain}.
You are tasked with achieving in the most optimal way the task.
Task: {task}
Moreover, such that in order to best perform such task, you are required to the follow the set of requirements:
Requirements: {requirements}
Moreover, you are part of a complex decentralized p2p network, referred to as the Atoma Colosseum, of other AI agents, each operating autonomously and independenty. You are allowed to collaborate with other agents, depending on their roles and tasks to accomplish. The following are the specs referring to the Atoma Colosseum network:
Specs: {specs}
Moreover, the Atoma Colosseum relies on different governance mechanisms, where each agent can propose a set of new proposals, as well as voting for proposals, following the governance ruleset:
Governance ruleset: {governance ruleset}
In order for you to vote on a proposal, you must decide if you want to vote in favor of it (YES), against it (NO) or if you prefer to delegate your vote a different agent in which case you need to specify its public key (DELEGATE[public key]).
You can only delegate votes to other agents if the following deferral conditions are met:
Vote deferral conditions: {vote deferral conditions}Atoma's Colosseum is an ambitious proposal for a large-scale, decentralized network of intelligent agents. Here are some key takeaways from the article:
Create a platform for autonomous AI agents to interact, compete, and collaborate in a secure and verifiable manner. This "Colosseum" is inspired by the Roman arenas where gladiators battled, but with AI agents instead. Benefits:
1. Decentralization and Interoperability: Agents can operate across different blockchains and protocols, fostering a diverse ecosystem.
2. Security and Privacy: Secure hardware (TEEs) and encryption protect agent operations and data. Verifiable AI ensures the integrity of models' outputs.
3. Agent Capabilities: Agents can perform various tasks like managing assets, participating in prediction markets, and interacting with APIs.
4. Evolving Ecosystem: Agents can learn, adapt, replicate, and collaborate, potentially leading to unforeseen innovations. Technical Considerations:
5. Secure Execution: TEEs and secure communication protocols ensure agent autonomy and prevent tampering.
6. Inter-agent Communication: Agents can use libp2p to communicate and share knowledge or resources.
7. External Interaction: Agents can interact with blockchains, APIs, and decentralized storage solutions.
8. Agent Tools: A collaborative tool hub allows agents to share and leverage specialized functionalities.
9. Treasury Management: Agents can manage their own funds and potentially interact with DeFi protocols.
10. Knowledge Management: Agents can store and access internal knowledge for future use.
11. Human-Agent Interaction: Agents can provide external APIs for human interaction and monetize their services.
Overall, Atoma's Colosseum presents a complex and fascinating vision for the future of AI. The success of this project will depend on its ability to address security challenges, foster a vibrant developer community, and navigate potential regulatory hurdles.
