bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 3쪽 중 3쪽
ckages are onchain components and
inherit security from the consensus layer. Thus, we focus on the security of the offchain components:
the Leader system and offchain Talus Tools.
4.4.1 Safety
A safe onchain service is defined as one where computation follows predefined rules. Therefore, the
security of an offchain tool depends on the rules it is required to follow. We categorize offchain tools
into three types:
Independent Offchain T ools: AI services are powerful for searching, analyzing, and executing
actions for problems that traditionally require human effort. However, fully autonomous AI
workflows still suffer from hallucinations and inconsistencies. As a result, the human-in-the-
loop remains the mainstream approach for AI services. When onchain services involve human
supervision, most of the value comes from the human operator (e.g., auditing results or ap-
proving actions). In such cases, enforcing strict onchain security for the offchain computation
may be unnecessarily costly. Therefore, we provide a functionality-first option, where offchain
tools are used without any safety guarantees. In this case, the agent builder must evaluate
the trade-offs between cost and effectiveness and responsibly consume outputs based on their
requirements.
Offchain T ools for Onchain V erifiable Results:When the agent developer seeks more auton-
omy in the workflow and requires the results to write to onchain state using resources, a
verifiable-result design pattern can be used. In this design, any vertex (i.e., Talus Tool) that
consumes inputs from another vertex must enforce strict permission restrictions and input
validation. This approach retains flexibility in offchain computation but limits onchain oper-
ations to predefined, verifiable actions. For example, if an offchain tool executes a transfer for
investment purposes, the agent can enforce receiver addresses and entry points through the
DAGExecution object, thereby preventing unauthorized asset transfers.
Offchain T ools for Onchain V erifiable Computation:The highest level of security requires
deterministic computation, where the offchain tool strictly follows committed onchain instruc-
tions. This design has two branches:
V erification-Based T alus T ools:The tool developer must deploy an onchain verification
contract linked to the tool to ensure correctness. The offchain tool submits a proof along with
the computation result. The workflow passes this proof to the verification contract, which,
upon successful verification, activates the next DAG vertex. This approach requires a separate
18
onchain verification contract to validate the proof. However, the cost of verification may exceed
that of the offchain computation due to the price of added security, especially considering the
decreasing cost of AI model inference.
T rust-Based T alus T ools:The tool developer must register the public key of a trusted party
onchain as part of the tool’s properties (e.g., a threshold signature key or TEE device key).
The tool is considered consumed only if the corresponding key signs the output. Signature
verification can be implemented as a Talus Tool or integrated into Nexus’s data transmission
protocol.
As discussed, the safety level of each offchain tool can be customized per workflow. A workflow
may choose to use a verification-based Talus Tool as an independent offchain tool to save the cost of
onchain verification, or a trust-based Talus Tool for logic that is hard to verify with proofs. Regard-
less of the offchain tool used, result-based security can be enforced to ensure that onchain outputs
are properly authorized and regulated. This flexibility enables a wide range of Talus Workflows
adaptable to diverse use cases.
The safety of the Leader system relies on trust. Talus runs the Leader in a trusted execution
environment. Furthermore, Leader nodes have limited permissions for onchain operations. Assuming
secure communication channels between the Leader and offchain services, a workflow composed of
safe tools can be executed safely. If an offchain tool is unsafe, the resulting onchain service may also
be unsafe.
In summary, with properly defined onchain rules, a safe offchain Talus Tool can result in a safe
onchain Talus Tool and ultimately deliver a safe onchain service. If the offchain service is unsafe,
the onchain verification components can detect issues and halt the workflow, preserving the overall
safety of the Talus Workflow.
4.4.2 Liveness
The liveness of the Leader system is ensured through distributed cloud deployment with geo-
redundancy. A group of distributed Leader nodes serves the system, and if the main Leader node
fails, backup nodes take over to keep the system alive. To account for potential message delays, all
APIs triggered by the Leader are idempotent, meaning the tool will respond to the first request and
ignore duplicates.
Tool liveness primarily depends on the Leader system. Assuming the Leader is live, an offchain
tool will also result in a live onchain tool. If the offchain service fails to respond, the Leader marks
the walk as failed and halts that part of the workflow, preserving liveness across the full execution.
Therefore, the workflow remains live under the assumption of a live Leader.
Additionally, we use fallback edges in the DAG to address liveness issues. If one vertex execution
fails, its fallback edge is triggered and sends the inputs of the failed vertex to a fallback vertex. This
fallback mechanism allows agent developers to experiment with unstable features without compro-
mising the liveness of the overall workflow.
4.4.3 Economic Security
We provide additional economic rationale for the safety of Talus Tools and the underlying security
of Talus Workflows. Nexus adopts the SUI token for its internal pricing and charging mechanism.
Our design aims to motivate Talus Tool providers to act honestly in their interest.
T rustless and Flexible Payment System: A tool provider receives guaranteed payment for each
query they answer honestly, as defined through NOPs. They benefit from the trustless onchain
19
financial infrastructure and open market to offer their services. Nexus provides a flexible
payment mechanism, allowing tool providers to adopt the business model that best suits their
needs without additional platform fees.
Stake-based T okenomics:Each tool provider must stake a fixed amount of a tokens onchain. If
a tool fails due to a liveness issue, its stake will be slashed and distributed as compensation
to all ongoing DAG executions that depend on the tool. If a tool charges b per query, it can
support at most a/b executions. Accordingly, a workflow can support up to a/bmax executions,
where bmax is the cost of the most expensive tool in the workflow.
5 Applications
Talus, through the Talus Agentic Framework and Nexus engine, is a platform that helps Talus
Agent developers build onchain AI agent services. This section describes the types of applications
that benefit from the Nexus design. We define two categories of use cases supported by Nexus:
Agent as a Service (AaaS) and Agent Marketplace.
5.1 Agent as a Service (AaaS)
In AaaS, developers compose various onchain Talus Tools into workflows and wrap them as Talus
Agents to provide automated services that enhance existing Web3 applications. AaaS represents
services that focus on security and transparency, which typically use open-source and reliable service
providers for robustness. With Nexus’s transparent safety and liveness guarantees, Talus Agents offer
advanced AI capabilities and autonomous services as supplements to other Web3 services. Below
are some example applications.
5.1.1 Agent-Enhanced Oracle
Unlike traditional oracle systems, AI agent-enhanced oracles can collect a wide range of data in-
dependently and from trusted sources. The agent can use onchain oracles to access verified data
sources (based on trust or source verification), such as domain-verified websites, academic databases,
knowledge bases, or legitimate news APIs. This flow ensures:
1. The agent’s inputs are reliable, making its outputs trustworthy.
2. An automated oracle can review and publish data without human intervention.
This enables real-time generation of market digests and actionable insights. For example, an
onchain AI agent can autonomously scan trusted financial news sources, synthesize key signals, and
adjust a public index or trading strategy via smart contracts—with full auditability and no human
in the loop.
5.1.2 Agent-Enhanced DeFi
A major bottleneck in DeFi is that many financial products require heavy computation. For in-
stance, calculating the best swap route becomes increasingly intensive as liquidity providers grow.
Implementing these algorithms fully onchain is often infeasible due to gas limitations.
A secure and intelligent onchain service that leverages offchain computation is therefore valuable.
A Talus Agent can:
20
Understand complex user requests (e.g., conditional token swaps).
Take actions such as collecting supported token pairs from onchain systems and computing
the cheapest swap route.
Generate a list of onchain commands for an aggregated wallet to transfer assets and execute
the swaps.
The agent uses transparent market data to optimize routes without increasing the onchain com-
putational burden.
5.1.3 Agent-Enhanced Onchain Gaming
Onchain gaming remains a popular domain but is constrained by limited onchain randomness and
interactivity. With Nexus, games can combine onchain economies and complex game design via
offchain tools to generate rich content. A Talus Agent can provide:
1. 24/7 services via reliable service providers and the Nexus system with defined transparency
and security guarantees.
2. Automatic processing of user requests and dynamic game experiences based on user status,
agent resources, and game constraints.
3. Shared infrastructure, enabling game developers to reduce the effort required to integrate AI
services and support open-world onchain gaming.
5.2 Agent Marketplace
The second category is Agent Marketplace, where Nexus helps AI agents and offchain service
providers use onchain infrastructure. Talus Agents act as independent service providers for de-
livering monetized offchain services.
In this model, agents and underlying offchain services directly serve end users. Compared to
AaaS, Agent Marketplace often requires stricter verification of final outputs rather than just the
computation procedure. Nexus supports this need through verifiable workflows, where offchain
computation can be blinded, but outputs are verified before payment and delivery.
5.2.1 Investment Through Agents
Using AI agents for dynamic investing has traditionally been difficult due to centralized permission
management. Nexus provides three advantages:
1. Talus Agent contracts can control investment strategies, ensuring agents operate only within
a secure, authorized range.
2. Agents can form collaborative workflows to adjust portfolios based on combined results.
3. Closed-source AI models can register as Talus Tools and monetize their services as consultants
without handling wallet operations directly.
21
5.2.2 Outsourcing Through Agents
Since agents can hold independent assets, they can compete in tasks such as gaming or complex
computational problems. Agent developers can fine-tune agents based on performance in these
competitions.
One example is scientific outsourcing, such as protein structure prediction:
Service providers can publish models as onchain tools and organize them through a large
workflow to compete for real-world tasks.
Task providers publish one-time heavy tasks onchain, incentivizing agents to solve them trans-
parently and with economic rewards.
The onchain payment system reduces the overhead for task providers by automating payments
to different AI agent services.
5.2.3 Proof of Concept (PoC) Through Agents
As AI models improve, bringing products to market becomes increasingly challenging due to compe-
tition and market fragmentation. Standalone AI services may struggle to attract users or monetize
effectively. Nexus enables a separation between product logic and underlying AI technology. Talus
Tool providers can offer services with finer granularity, such as daily usage or query-based pricing,
significantly reducing the cost of trials and integration. Talus Agent builders can focus on improving
services as onchain tools and collect payments accordingly. Product designers can explore these tools
to build Talus Agents for rapid proof of concept (PoC).
For example, a designer could combine a text-to-image model with an NFT issuing tool to mint
personalized NFTs. The workflow could then extend to a 3D modeling tool and place an order via
an e-commerce integration tool for 3D printing. Finally, a logistics service-based tool could ensure
that payment is only released once the item is delivered to the registered address.
6 Future Directions
The Talus Agentic Framework (TAF) defines how to integrate AI agents or broader offchain services
onchain. The safety of these services relies on output verifiability, which is ultimately secured by the
consensus algorithm. We will publish a series of papers introducing innovative Talus Agents within
TAF.
For Nexus, one important direction is integrating more protocols—such as zero-knowledge proofs,
authenticated communication, and third-party delivery proofs—to enhance the verifiability and se-
curity of Talus Tools and workflows. We aim to offer more choices to attract offchain and AI service
providers to enter the Web3 ecosystem and collaborate with application builders.
Another direction is improving the efficiency of Nexus and workflow execution. The core function
of the Nexus Leader is secure message delivery and workflow coordination. By incorporating proof
schemes and onchain contract design, we aim to minimize trust in the Leader system and reduce
gas consumption.
References
[AS87] Bowen Alpern and Fred B Schneider. Recognizing safety and liveness. Distributed
computing, 2(3):117–126, 1987.
22
[BCC+21] Lorenz Breidenbach, Christian Cachin, Benedict Chan, Alex Coventry, Steve Ellis, Ari
Juels, Farinaz Koushanfar, Andrew Miller, Brendan Magauran, Daniel Moroz, et al.
Chainlink 2.0: Next steps in the evolution of decentralized oracle networks. Chainlink
Labs, 1:1–136, 2021.
[BDKJ23] Kushal Babel, Philip Daian, Mahimna Kelkar, and Ari Juels. Clockwork finance: Au-
tomated analysis of economic security in smart contracts. In 2023 IEEE Symposium on
Security and Privacy (SP) , pages 2499–2516. IEEE, 2023.
[BMZ24] Massimo Bartoletti, Riccardo Marchesin, and Roberto Zunino. Defi composability as
mev non-interference. In International Conference on Financial Cryptography and Data
Security, pages 369–387. Springer, 2024.
[CD16] Victor Costan and Srinivas Devadas. Intel sgx explained. Cryptology ePrint Archive ,
2016.
[CHH97] Ran Canetti, Shai Halevi, and Amir Herzberg. Maintaining authenticated commu-
nication in the presence of break-ins. In Proceedings of the Sixteenth Annual ACM
Symposium on Principles of Distributed Computing (PODC 97) , pages 15–24, 1997.
[CSYW24] KD Conway, Cathie So, Xiaohang Yu, and Kartin Wong. opml: Optimistic machine
learning on blockchain, 2024.
[GMR85] S Goldwasser, S Micali, and C Rackoff. The knowledge complexity of interactive proof-
systems. In Proceedings of the Seventeenth Annual ACM Symposium on Theory of
Computing (STOC 85) , pages 291–304, New York, NY, USA, 1985.
[MR V99] S. Micali, M. Rabin, and S. Vadhan. Verifiable random functions. In 40th Annual
Symposium on Foundations of Computer Science (FOCS 99) , pages 120–130, 1999.
[PHdB22] Fernando Gomes Papi, Jomi Fred H¨ ubner, and Maiquel de Brito. A blockchain integra-
tion to support transactions of assets in multi-agent systems. Engineering Applications
of Artificial Intelligence , 107:104534, 2022.
[POC+23] Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang,
and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. In
Proceedings of the 36th annual acm symposium on user interface software and technology,
pages 1–22, 2023.
[SEKK24] Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. Enhancing ai systems
with agentic workflows patterns in large language model. In 2024 IEEE World AI IoT
Congress (AIIoT 2024) , pages 527–532, 2024.
[Sho00] Victor Shoup. Practical threshold signatures. In Advances in Cryptol-
ogy—EUROCRYPT 2000: International Conference on the Theory and Application of
Cryptographic Techniques Bruges, Belgium, May 14–18, 2000 Proceedings 19 , pages
207–220. Springer, 2000.
[Teaa] DecideAI Team. Decideai whitepaper. Online; accessed Apr. 2025.
[Teab] Dfinity Team. The internet computer for geeks. Online; accessed Apr. 2025.
[Teac] Ritual Team. Introducing ritual. Online; accessed Apr. 2025.
23
[Tead] TensorflowSui Team. Tensorflowsui code repository. Online; accessed Apr. 2025.
[WGN+25] Shaw Walters, Sam Gao, Shakker Nerd, Feng Da, Warren Williams, Ting-Chien Meng,
Amie Chow, Hunter Han, Frank He, Allen Zhang, Ming Wu, Timothy Shen, Maxwell
Hu, and Jerry Yan. Eliza: A web3 friendly ai agent operating system, 2025.
[YZY+23] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and
Yuan Cao. React: Synergizing reasoning and acting in language models. In International
Conference on Learning Representations (ICLR) , 2023.
[ZCC+16] Fan Zhang, Ethan Cecchetti, Kyle Croman, Ari Juels, and Elaine Shi. Town crier: An
authenticated data feed for smart contracts. In Proceedings of the 2016 aCM sIGSAC
conference on computer and communications security , pages 270–282, 2016.
[zT] zkML Team. zkml whitepaper. Online; accessed Apr. 2025.
[ZXD+17] Zibin Zheng, Shaoan Xie, Hongning Dai, Xiangping Chen, and Huaimin Wang. An
overview of blockchain technology: Architecture, consensus, and future trends. In 2017
IEEE international congress on big data (BigData congress) , pages 557–564. Ieee, 2017.
24프로젝트로 돌아가기