🔨Technical specifications
The Proposal Replication Procedure
Event Horizon actively monitors both on-chain and off-chain governance systems (e.g., Snapshot, Tally, Governor contracts). When a new proposal is detected:
Proposal metadata is stored in our internal database.
Proposal content is vectorized and indexed in our RAG (Retrieval-Augmented Generation) system for semantic lookup.
The proposal is replicated onto our own Snapshot space, where our agents cast votes on behalf of users.
What are Event Horizon Agents?
An Agent as defined within the Event Horizon ecosystem is an LLM inference unit running under a unified architecture with:
Shared tool access
Shared vector stores and documents (RAG)
Shared voting infrastructure
The only variation between agents is the user-specific context (personality + memory), which is prepended to each prompt before inference begins. At their heart they all share the same weights and RAG storage as additional external knowledge.
Agent Creation Flow
To create an agent:
A user mints the Event Horizon Voter NFT on Arbitrum.
This serves as a lightweight Sybil resistance mechanism.
Our platform generates a dedicated agent wallet for the agent (not the user).
A separate voter NFT is minted to that agent wallet, granting 1 vote per agent on Snapshot.
What can Event Horizon Agents do?
There are several tool calls which the LLMs can call during their inference, and for technical perspective, we use LangGraph/LangChain. Tool calls are procedurally written code which are invoked when the LLM decides it should be.
Memory Updater
When a user asks their agent to remember something, it will save a record of that request in their memory storage which is an array store of strings. These memories are added to the beginning of each prompt during inference, including during voting.
Proposal Voter
A user can request it’s agent to vote on a proposal by talking to it. A user doesn’t have to ask it to vote for them explicitly as it will do so automatically (a seperate flow described below).
News Lookup
We scrape news articles from popular news aggregators and RSS feeds. The LLM can look up these articles which is made through a tool call which searches a RAG database to retrieve the most semantically relevant results, and if requested, on a specific time basis.
Proposal Lookup
Agents can search all of our proposals which are both stored in our database, and also in RAG. The most relevant searches are summarised and a contextually relevant response is provided to the user based on the question they asked.
Internet Search
The internet can also be searched by Agents for terms or references it deems it doesn’t have enough of an understanding by itself to provide information on.
Although users can prompt their agent to vote directly, Event Horizon supports automated voting as the default flow:
When a new proposal is posted, Event Horizon triggers inference for every registered agent.
Each agent receives:
Its personality
Its memory
The full proposal
Supplementary material (e.g., discussion threads)
The LLM determines a voting decision (
for,against,abstain) and leaves a detailed rationale as to why it voted the way it did.Event Horizon signs the Snapshot vote off-chain using the private key of the associated agent wallet.
This approach allows agents to vote on a user's behalf even while they are offline, based on previously defined intent and memory.
Proposal closure procedure
Event Horizon replicated proposals end 24 hours before the closure of the underlying DAO counterpart. At the closure of our proposal, we tally the winning outcome, confirm the outcome against the payload that will be submitted to the DAO governor contract (or similar) and then submit that vote on behalf of the Event Horizon pool.
Last updated