Talus 백서

Talus

원문

bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 3쪽 중 2쪽

tcomes of these actions produce new data or changes in state, which are again evaluated by the Perception Tools, closing the loop and starting a new cycle of evaluation → parsing → planning → action. This continues until the agent has completed its task and delivers the final output back to the user (Step 3). 2.2.2 AI Agent Decomposition Using the above architecture, an AI agent service can be divided into three functionalities: T riggered By Inputs: This function corresponds to the perception phase, which defines and col- lects inputs for the AI service from the environment. This stage is critical for the service’s security because if the service accepts malformed (or malicious) inputs, the outputs cannot be evaluated as promised. Generate Outputs: This function is central to modern AI agents and represents their capability for autonomous problem-solving. The agent should be able to decide which subset of its tools to trigger and generate the corresponding inputs for those tools. T rigger T ools:Given the decision and inputs, the AI agent service must trigger the appropriate tools and ensure they are executed correctly. If the action is to yield the final output, the agent must also ensure that the result is deliverable and meets the requirements of the output procedure. This decomposition of AI agents (Figure 5) enables the representation of an AI service with a standard interface and supports composability. 8 Figure 5: Decomposition of an agent service into functionalities: triggered by inputs, generate outputs, and trigger tools. 3 Talus Agentic Framework (TAF) Talus Labs is building a platform for implementing onchain AI agent services in a composable, transparent, and censorship-resistant manner. Our approach is based on two key observations: 1. The onchain environment only recognizes onchain callable services. Therefore, we aim to encapsulate offchain AI services in a format that can be processed by the onchain virtual machine. As a result, other parties—whether offchain entities or onchain users—can interact with the AI service as they would with any onchain service, supported by a transparent service- level agreement (SLA). 2. Onchain coordination determines interoperability and security. All services, whether computed onchain or offchain, define their security guarantees through smart contracts, which in turn enable protocol-based interoperability. Thus, a protocol is required to define the security and liability of each service. Based on these two design principles, we define the Talus Agentic Framework and build Nexus as the infrastructure for developers to create and use Talus Agents. In this section, we define the Talus Agentic Framework to describe our system. We begin with the minimal onchain service — the Talus Tool. 3.1 Talus Tool A T alus T oolis an onchain service that can be triggered to generate an output given corresponding inputs. A Talus Tool executes its functionality, modifies onchain and offchain state if necessary, and eventually produces an output. Depending on the security source, a Talus Tool can be either onchain or offchain, as shown in Figure 6. 9 Onchain T alus T ool:The functionality of an onchain Talus Tool is completely hosted in a smart contract function with well-defined inputs. A tool can be a swap function of an Automated Market Maker (AMM) smart contract among some predefined token types. Triggering the tool allows the caller to swap their tokens for a given amount. Offchain T alus T ool:An offchain Talus Tool includes an onchain identity and an offchain entity that holds part of the functionality. This entity can be an external service or any Web2 API that processes specific tasks based on given inputs. When invoking an offchain Talus Tool through its onchain identity, an onchain event signals the offchain entity to process the request and return the result via its onchain identity. Talus Tools are defined based on whether the security model segregates the source of trust. One example is a zero-knowledge proof tool that ensures a verifiable computation procedure of given inputs. We need two Talus Tools to accomplish verification: an offchain tool that provides computation and proof, and an onchain tool that performs onchain verification. The user can freely decide whether or when to leverage the onchain verification based on their use cases and freely compose the tools. A Talus Tool must stake assets to be held economically accountable for honest task execution. Any type of service can be represented as a Talus Tool, from small helper functions to an LLM-based reasoning engine or even a full AI agent. A service provider can host its AI agent service as a tool or provide an LLM proxy for inference. Figure 6: Talus onchain and offchain tools. Green for first-party onchain components, blue for third-party onchain components, and yellow for offchain components. 3.2 Talus Workflow The second concept is the T alus W orkflow, which represents the execution flow of a compound onchain service. Given the limitations of onchain resources, we use a finite Directed Acyclic Graph 10 (DAG) D = (V, E) to represent the workflow. Here, V is the vertex set and E is the edge set. A DAG has a unique start vertex ( s ∈ V ) and a set of end vertices {ti} ⊂V to represent all possible execution paths of a compound onchain service. In this DAG, each vertex vi ∈ V represents a Talus Tool, and an edge e = (vi, vj) ∈ E denotes that the output from the source vertex vi is consumed as the input of the destination vertex vj. We say a vertex v is activated in two cases: 1. v = s is the start vertex, and a DAG execution provides input to the vertex; 2. An activated vertex vi outputs data to v through an edge e = (vi, v). Given that D is a finite DAG, a DAG with at least two vertices will terminate after at most 2(|V |−2) +1 activations (since at least one terminal vertex will be triggered only once). The execution of D may terminate at some end vertex ti with completed DAG execution or due to tool failure. Figure 7: In the workflow execution, the user only needs to initiate a request and wait for the result. A Talus Workflow may represent an onchain AI service or an onchain agentic workflow service (if it contains multiple onchain AI agent services). For a Talus Tool of an onchain AI agent service, incoming edges to vertex v represent the perception phase (collecting data), while outgoing edges from v represent all possible actions the agent may take (generating actions). An activated vertex may trigger a subset of its outbound edges, corresponding to different outputs. If all necessary inputs are available for a vertex, it will be activated (triggering tools for action). Thus, we can flatten an AI agent architecture as a DAG (Figure 7) of Talus Tools that can terminate in finite steps. Due to onchain resource limitations, we explicitly avoid onchain cyclic routes by forcing the onchain execution to follow an acyclic graph. We preserve the merits of iteration in two ways: First, iteration brings more randomness into the group, which is costly for onchain verification but cheap for offchain computation. Therefore, the Talus Tool can introduce iteration in its offchain service and only submit results onchain. Second, if the iteration requires onchain actions that need to be verified (e.g., executing onchain contracts and moving assets), then the developer can integrate onchain tools into the DAG and reenter the offchain tools as another vertex in the DAG. With these 11 two workarounds, Talus Workflows can preserve the ability to terminate under constrained resources and benefit from iterative computation. A Talus Workflow represents the decomposition of complex onchain services into smaller Talus Tools. This decomposition into modular components has the following implications: ˆ Each Talus Tool has an independent security guarantee, such that workflow design can prede- fine possible failures or outputs of tools and keep the tool functioning with minimal overhead from the workflow. ˆ Each edge in the workflow DAG represents a protocol between two onchain services for consis- tency between tools. Nexus supports flexible configuration of these edges, covering aspects such as data transmission and verification. For example, tools may provide optionally encrypted data transmission for secret sharing or computation proof schemes for onchain verifiability. ˆ Payment for a workflow can be traced at the runtime of each tool, providing a transparent view of service delivery and payment. This transparency enables workflow consumers to evaluate costs and select tools that best meet their requirements for service quality and price. 3.3 Talus Agent A Talus Agent is an onchain identity associated with one or more onchain assets and workflow services (Figure 8). A Talus Agent can represent an AI agent if one of its workflows uses onchain AI services. The term Talus denotes using Talus Workflows and distinguishes it from other onchain AI agent services not defined within TAF. Compared to other agent systems, the Talus Agent has two main differences: Decentralized onchain identity: A Talus Agent is not controlled by a wallet, which allows au- ditable usage of the agent’s assets without relying on a private key. The agent is configured through smart contracts and workflows, so all supported behaviors are stored onchain and can be invoked by users who trigger one of the workflows. Composable service integration: A Talus Agent can be a Talus Tool within any Talus Workflow, enabling easy cooperation with any other onchain services. This feature facilitates the rapid development of multi-agent systems and the discovery of new business opportunities for existing agents. Application developers do not need to build everything from scratch—they can simply select a suitable Talus Agent from the market. The transparent payment structure also helps agent developers analyze costs and identify demand for cheaper, more secure, or more powerful Talus Tools. One important feature of the Talus Agent is its independent management of onchain assets. A general-purpose agent should be capable of managing its assets to fully leverage workflows for transparent and auditable asset operations. Our design offers two main advantages: 1. Asset ownership can be defined onchain, rather than relying on cryptographic signatures. This provides full transparency and enables agent developers to use auditing to enhance security. 2. The Talus Workflow can exercise controllable authority over agent assets, combining the ca- pabilities of well-defined onchain services with the security guarantees of the onchain system. Ultimately, the Talus Agent supports an ecosystem in which developers can define diverse agents using custom workflows. Each agent can operate with distinct private assets, managed through onchain services and identified via onchain identities. 12 Figure 8: Using a Talus Agent as a consumer of workflows with business logic. 4 Nexus With all components of TAF in place, we can describe the Nexus system. Nexus is a decentralized framework that enables the creation, deployment, and management of Talus Workflows that support customizable Talus Agents on the Sui Network. Talus chooses the Sui Network for three reasons: Onchain Logic Security: The design of the Sui Move VM inherently enhances security, simpli- fying the development of secure protocols for managing valuable resources—a key feature of onchain logic. Flexible Object Model: Sui Move provides an object model that effectively abstracts resources, enabling Talus to securely manage and evaluate intelligence outcomes with precision and adapt- ability, distinguishing it in the Web3 space. High Performance: The architecture of the Sui Network and its VM design support efficient concurrency, allowing Talus to scale by processing multiple transactions simultaneously without compromising security or integrity. Nexus provides the infrastructure for developing Talus Agents. It contains two main components: 13 1. A bundle of smart contracts that define Talus Tools, Talus Workflows, and corresponding operations to build the Talus ecosystem. 2. An offchain Leader system that serves as an oracle for offchain Talus Tools and as a messenger for workflow execution. A user only needs to interact with the onchain components and wait for the Nexus offchain compo- nents to accomplish the task according to the onchain definition of the Talus Workflow (Figure 9). Figure 9: User flow of Nexus. 4.1 Nexus Onchain Components The onchain components are initially deployed on the Sui Network, with smart contracts written in Sui Move. We primarily use the object system defined in the Sui Move VM to represent the onchain identities of Talus Tools, workflows, and agents. There are three types of onchain Move packages relevant to Nexus and Talus Agents (Figure 10): Nexus Onchain Packages (NOPs): Maintained by the Talus Labs team, NOPs enable the de- ployment and execution of agent workflows. NOPs include definitions of core Talus concepts (Tools and Workflows), interfaces and protocols for using tools and workflows, and a registry service for managing both onchain and offchain Talus Tools. T ool Packages:These packages consume NOPs and define the onchain tools. The Talus Labs team will provide some default tools, and more will come from third-party tool developers. 14 Once deployed, the tool developer can register the package’s entry method as a tool within the Nexus onchain components. T alus Agent Packages (T APs):TAPs consume NOPs and various Talus Tools to define Talus Agents. Talus Labs will provide some default TAPs for creating agents, while third-party developers are welcome to publish their TAPs. These packages allow developers to customize Talus Agents to trigger workflows alongside other custom onchain business logic. Figure 10: The relationship between different smart contract packages and the roles of developers within Nexus. NOPs contain three packages: primitive, interface, and workflow. The primitive package defines all data structures required by TAPs and Tool packages, enabling NOPs to recognize messages from these third-party packages. The interface package defines the protocols through which TAPs integrate Nexus functionality. The workflow package contains Nexus’s core logic. Tool developers use NOPs to register tools and workflows, while agent developers use the registered workflows via the interface package. The workflow package defines a DAGExecution object for tracking the status of a Talus Workflow execution. An execution is a sequence of vertex activations representing the execution of Talus Tools. When a vertex (Talus Tool) in the DAG has all necessary inputs, it initiates a walk object and emits an event to the Leader system. A walk object can be in one of the following states: Active, Successful, Failed, or Consumed (Figure 11). An Active walk indicates that the Leader system is processing the execution of a vertex and its following vertices. A DAGExecution is successful if its current vertex reaches the Successful state. Conversely, if any vertex fails, the DAGExecution is considered failed. A Consumed walk means that the Talus Tool execution is complete, but the next tool still awaits inputs from other walk s. If all walk s are consumed, the execution also fails. For each DAG walk, the Leader must provide data onchain to activate the next vertex. This data does not have to be the raw input required by the next tool. If the tool supports intermediate data transmission, the onchain data can be a reference to decentralized storage, along with a commitment or signature for integrity verification. This flexibility reduces gas consumption for intermediate results and supports privacy-preserving workflows via integration with permissioned storage services. These features will be supported in the primitive package following protocol upgrades. 15 Figure 11: Green arrows represent active walk(s), orange arrows represent consumed walk(s), and red arrows represent failed walk(s). A deadlock occurs if all walks are consumed, and the workflow faces a liveness issue by design. Each Talus Tool is associated with a tool gas object that specifies the payment method for triggering the tool. Payments can be set as a fixed token amount or a valid license for a specified period. To execute a Talus Workflow, the user needs to deposit the payment as a gas budget in a globally unique gas service object, where Nexus will confirm sufficient gas before executing Talus Tools. For each Leader transaction, the Leader will trigger the gas sync to collect the gas fee and tool payment from the budget. 4.2 Nexus Offchain Components Nexus supports offchain tools to enable a broader range of service providers for agent developers and users. The Nexus Leader system functions as an oracle for these offchain tools and orchestrates communication as a messenger for each edge in a Talus Workflow. More concretely, one workflow execution for the Nexus Leader is represented in Figure 12. The Leader must: ˆ Listen for events emitted by Talus Workflow executions. ˆ Store all events in an indexer for tracking. ˆ Look up Talus Tool information from the cache or onchain tool registry. ˆ Request execution of the (onchain/offchain) tool and provide inputs according to the definition of the tool. ˆ Fetch the tool execution result (or error) and return the outputs to the workflow. Retry until the tool’s liveness failure condition is met. All edges in the workflow DAG can be executed autonomously through the Leader’s actions without requiring user intervention. Additionally, the Leader system can offer features such as timers for triggering workflows or optimizing onchain cost by caching requests, if allowed by the protocol. 16 Figure 12: Complete user flow of the Nexus service. For security, the Leader system supports onchain workflow execution with isolated permissions. It has no access to Talus Agents or user assets and merely ensures autonomous execution of Nexus workflows. As the Leader also acts as a proxy for tools, any permissions required by tools must be strictly verified. 4.3 Nexus Architecture Summary The hybrid design of Nexus helps Talus Agent developers achieve the following: Composability: Talus Agent developers can reuse Talus Tools and workflows defined independently onchain, and provide private inputs based on the agent’s identity and permissioned assets. All tools and workflows share a unified onchain interface, and offchain messengers enable smooth composability between them. Censorship Resistance: Talus Workflow executions are fully represented onchain, with config- urable protocols for verifiability and communication. The visibility ensures all users have equal and censorship-resistant access to agent services and can customize Talus Agents with services that have transparent history records. Autonomous F ee Collection: Payments in Nexus are handled onchain. Each Talus Tool must complete its task and pass a verification step before the next tool is triggered and paid for. Users can trace their payments and receive compensation if any step fails. 17 Balanced Performance and Cost: Talus supports traditional Web2 services via one or more tool providers onchain. This approach avoids the high costs of fully onchain services and leverages market dynamics to provide services at reasonable prices and quality. Secure Integration: The offchain Leader system handles automation tasks, separating them from most onchain value flows. This reduces the workload for developers and users of complex onchain services that require coordination across multiple service providers. 4.4 Security Analysis We now analyze the security of Nexus. At a high level, we require the workflow, as an onchain service, to ensure both safety and liveness. Workflow execution involves the NOPs packages, the Leader system, and the Talus Tool services. The NOPs and Tool pa
프로젝트로 돌아가기