The Architecture

The System Architecture

The Web3 coordination layer of agent/acc is designed to anchor agent configuration, behavioral governance, staking dynamics, and discoverability to-chain. This complements the offchain AI stack — including data ingestion, fine-tuning, and operations — by ensuring trustless integrity, public accountability, and economic incentives. While core AI infrastructure (e.g., ingestion, training) runs offchain for performance, key control mechanisms are enforced via onchain contracts deployed on Base, an Ethereum Layer 2 network, to ensure optimal efficiency, security, and scalability.

The current implementation of agent/acc operates on a modular onchain/offchain architecture. While $A/ACC is deployed for liquidity on Uniswap V3, the following smart contracts are in development to enable core functionalities:

AgentRegistry.sol

• Stores agent metadata (identity hash, config pointer, LLM signature)

• Allows creation, update, or forking of agents with associated fees

• Emits events for offchain indexing (new agent created, config updated)

StakingController.sol

• Allows $A/ACC to be staked against registered agents

• Calculates and exposes performance-based weights

• Facilitates slashing based on reputation or protocol-defined heuristics

AgentGovernor.sol

• Proposals: Agent config updates, behavioral shifts, tone or scope changes

• Voting: Token-weighted or delegation-based vote casting

• Execution: Pushes updates to AgentRegistry on majority approval

These contracts will be deployed in stages post-audit. In the meantime, the protocol utilizes offchain verification and whitelisted relayers to simulate these behaviors within the private beta.

Onchain ↔ Offchain Integration

agent/acc operates with a hybrid architecture where smart contracts enforce ownership, reputation, and behavioral control over agents, while most computationally intensive operations (data ingestion, fine-tuning, memory recall) happen offchain.

Component	Role	Onchain Interface
Data Ingestion Engine	Parses and preprocesses user data	Authenticated ZIP submission hash is recorded onchain
Model Fine-Tuning	Generates LoRA adapters and stylistic embeddings	Embedding metadata (e.g., LoRA checksum) is committed via AgentRegistry
Memory State	Dynamic context store (vector index)	IPFS/Arweave CID registered in config slot
Interaction Engine	Composes tweets, replies, posts	Interactions signed by agent keys (verifiable offchain)
Governance Layer	Proposals to overwrite memory, fork agent, update tone	Enforced via AgentGovernor and Snapshot
Performance Analytics	Tracks engagement, uptime, prediction accuracy	Offchain scoring, slashing recorded via StakingController