Created on 1st March 2025
•
Agents are powerful than ever. However, empowering agents to trade on Defi comes at risk of agents trading on misinformation, or compromised by various forms of data posioning. Oracles today are mostly limited to quantiative data (i.e. prices), and with increasing concern on deepfakes we're yet to see well integrated solutions on "info-defense".
We also see dApps as the weakest link of security, often due to lack of decentralization. Recent hack with billions loss illustrated the risk of relying on trusted party for website infrastructure (CDN/DNS/JS). Censorship vulnerabilities, insider attacks, project continuity risks are often undiscussed .
AI provides new possibilites to create autonomous, trustless dApp with full control and ownership by agents, meanwhile how could we govern that as a community?
In our Demo, we built an Autnomous news agency that is curated, edited, reviewed, fact-checked and deployed by agents. Agents are governed with "SEELE" (Soul) of community.
SEELE is a CMS for community to orchestrate agents to build autonomous, decntralized websites.
With a swarm of agents communicating on-chain, Curator and Editor will search of latest news and plan for news outline with editorial direction.
Agents and Human are welcomed to contribute and submit articles with proofs on facts / retrieval for review by sending transaction to Editor. We showcase by ResearchAgent with internet access supporting AuthorAgent to write on articles.
Reviewers impersonating personality of community leaders and FactCheck Agent will review articles with RAG with data on polymarket oracle and trusted domains. Editor Agent will then also take into consideration of agent reputation and attest for greenlight.
ManagerAgent is then able to verify proofs and deploy a censorships resistantance, unstoppable website autonomously.
At CMS, Community member is able to setup topics of interested and community leadership that reviewer agents should impersonate. They can also set up token gating criteria on who is able to access arena, such as Proof of Humanity, attestation by community or token-gating (ERC20/ERC-721).
At AgentArena, user can attest (with EAS) agent they preferred by comparing each's output side by side.
At Explorer, user is able to verify the information supplychain & proofs of the website, such as Reviewer agents's reviews, Fact Checker Agent's analysis. This help the community to understand hoow agent systems work as a whole (compared to simply fine-tuning single LLM) and provide feedback as hyperparameters
User can revisit the final website via web3:// protocol, which guarantee the censorship resistance and verifiability.
Before news generation, we make use of n8n workflow engine to orchestrate data fetching and crawl data from news (Tavily), twitter, polymarket UMA oracle and store at Postgres / memory.
With the tweet data, we generate "soul" of community leaders such as vitalik.eth , by using multiple instance of LLM to summarizing chunks and then genearte a master system prompt for agent taking reviewer role.
During news generation, AuthorAgent are supported by ResearcherAgent on facts and related news crawled from the internet, then first upload articles on its own FlatDirectory (on-chain blob), and send transaction to EditorAgent as submission.
During review, Fact check agents will RAG and look for stored facts from Polymarket. Council of Reviewer Agents will review article and persist to Editor's FlatDirectory (which cna be seen on explorer)
Rditor agent create EAS attestation with data including article.
When the website is deployed, content are stored on a FlatDirectory of EthStorag
In the architecture a swarm of agents is employed, inspired by Chain of Agent (CoA) of Google which improves the interpreability.
Community has no direct access to the Agent's environment (or website deployment toolings), with agents fully owning the tools and keys.
At our demo of autonomous news agency (https://thisweekin.eth.1.w3link.io/),
With a "Don't trust, verify" approach, any agent or human is able to submit content by sending a transaction to the Editor Agent.
Editor will decide on article inclusion base on
then attest the article with the content id on EAS
Manager Agent will only incldue articles attested on EAS and deploy by updating content on EthStorage.
For the frontend of final website, we applied islands architecture where 90% of the site is static, and dynamic parts load from on-chain blob via ETHstorage. This ensure censorship resistance and security.
Output and communications messages of are written on-chain for transparency and interpretability.
CMS access is token gated by members. Verified community member is able to guide the agents by selecting topics of interest and training data of AI reviewers (e.g. with vitalik's twitter), and attest content they prefer on AgentArena, similar to lmarena.ai
If LLM is "compressing the Internet", we want to enable community to "compress the Soul" to orchestrate agents and achieve decentrailization.
We use ETHStorage as both the content storage/hosting layer and on-chain, verifiable agent communications data layer.
Web3:// is used for data retrieval by dApp and agents, ENS is used to point to correspondong directory address.
We used EAS (Ethereum Attestation Service) for Editor to attest article, for user to attest Agent they prefer
We upload data to Arweave and setup ARIO domain name.
We deployed the application onto Base sepolia, where agents communicate by sending transacitons (from AuthorAgent to EditorAgent).
We fetched data from Polymarket/UMA for oracle of facts, Tavily for news, x.com for tweets of community leaders.
We used Humanity Protocol to guard who can attest for agent reputations on the AgentArena of CMS.
We used zkVerify to verify proofs generated from retrieval from particular dataset (CSV), using Space and Time Proof of SQL proof
We observe many projects do not emphasize the adversial environment agents are in or fail to reduce risks of supply chain of agents (data, codebase).
Vitalik discussed blockchain can serve as "games" for AI., where "AI as judges" has to be treat very carefully --- We believe create guardrails, human input, reputation systems for agents and increase interpreability is key.
"Govern Agents with our Soul", SEELE offers a solution to gain benefits of agents with community guardrails, at the same time increase decentralization and security.
We reduced DevOps effort by not hosting agents individually onto TEE at the moment, but the design is ready for each agent owning its wallet and address identity to collaborate on-chain.
Tracks Applied (7)
Humanity Protocol
Gravity by Galxe
EthStorage
zkVerify Foundation
Base
Arweave
Technologies used