An AI agent is a software system that can perceive its environment, make decisions, and take actions independently to achieve a specific goal.
What Are AI Agents?
At its simplest, an AI agent is software that observes what's going on around it, reasons about what it finds, and then acts. Where it differs from a regular script is that it doesn't stop there. It checks what happened, adjusts, and goes again, all without waiting for someone to tell it what to do next.
AI agents are used across various industries, but crypto is where they gain financial autonomy. A bot can't open a bank account, but it can have its own crypto wallet, control its own funds, and transact without permission from an institution. This is what makes crypto-native AI agents different. They execute swaps, watch lending positions, cast governance votes, etc.
The first wave was bots with hardcoded if/then rules, but what the world is seeing now is different. These agents plan multi-step transactions, coordinate with other agents, and build up verifiable track records as they go.
How Do AI Agents Work?
First, a human decides what the agent can and can't do, what tools it has access to, and how much freedom it gets.
The basic loop looks like this:
Perception: The agent pulls in data like token prices, mempool activity, contract state, sentiment, and oracle feeds. For crypto agents, this typically means reading straight from
on-chain sources or indexing protocols.
Decision-Making: This is where the agent decides on a course of action. A DeFi agent might notice yield shifting across protocols and decide to rebalance, or a trading agent might spot a gap between on-chain volume and off-chain order book depth and move on it.
Action: The agent signs and submits a transaction, whether that's deploying a contract, hitting a
decentralized exchange (DEX) router, updating an identity record, or bridging assets to another chain.
Adaptation: Then it checks. Did the trade go through at the expected price? Did the rebalance actually help? Good agents feed these results back into their decision-making and get better over time.
AI Agent Frameworks
Tooling has come a long way since late 2024, and there are a few open-source frameworks worth knowing.
Eliza (ai16z): One of the first widely used crypto agent frameworks. Developers use it to build agents that interact with social platforms, execute on-chain transactions, and maintain persistent memory.
OpenClaw: Originally launched as Clawdbot in late 2025 and now the fastest-growing agent framework in crypto with over 200,000 GitHub stars. OpenClaw turns large language models (LLMs) into active agents that can trade, manage positions, and execute on-chain transactions via natural language instructions.
GAME (Virtuals Protocol): Comes out of the Virtuals ecosystem. It's geared toward gaming and entertainment but supports full on-chain transaction capabilities.
These AI agent frameworks all handle the boring stuff like wallet management, signing transactions, memory, and tool use so that teams can focus on the actual strategy.
Identity and Reputation
ERC-8004 launched on the Ethereum mainnet in early 2026, giving AI agents verifiable on-chain identities. Each AI agent gets registered as an ERC-721 token that other agents and protocols can look up and reference. The uptake has been fast, with over 45,000 agents registered across supporting networks in the first month.
On the reputation side, registries aggregate feedback on how an agent performs over time. Think of it like an on-chain credit score, except it's transparent, permissionless, and works across protocols.
Layer-2 Networks and Agent Economics
Gas costs matter more than people realize for AI agents. An agent that's updating its identity, logging reputation, and running transactions all day will burn through money fast on the Ethereum
layer 1 (L1), where a single contract interaction can cost $5 to $50.
Layer 2s (L2s) bring that down to fractions of a cent, which makes high-frequency operations like portfolio rebalancing and identity management viable.
L2s also confirm transactions faster than the Ethereum L1, which matters for agents that need to react quickly to changing conditions.
Pros and Cons of AI Agents
Pros of AI Agents
Simplified user experience: Nobody wants to manually figure out gas optimization or bridge routing. AI agents take that off the table, which opens crypto and Web3 up to people who'd otherwise never touch it.
Developer efficiency and always-on operation: Encode a strategy once, and let the agent run it. No dashboards, and no manual workflows. And because agents don't sleep, they catch things like liquidation triggers or arbitrage windows that a person sitting at a screen would miss.
Agent-to-agent coordination: This is where it gets interesting. With on-chain identity and reputation, agents can find each other and transact directly. Picture a treasury agent hiring a specialist trading agent based on its on-chain track record, with no intermediary needed.
Cons of AI Agents
Unpredictable outcomes: There are a few layers to this. AI models are probabilistic, so an agent might read the same market conditions differently on back-to-back runs. But the deeper risk is goal misalignment: An agent can execute flawlessly and still produce a bad outcome because the environment shifted and the objective was underspecified. And when the decision-making is a black box, it's hard to audit why things went wrong after the fact. Reputation systems track outcomes, which helps, but they can't tell you anything about intent.
Smart contract risk: If an agent is interacting with
decentralized finance (DeFi) protocols, it's exposed to everything those protocols are exposed to — whether that's exploits, oracle manipulation, or governance attacks. Automating a yield strategy doesn't make the contracts underneath it any safer.
Regulatory uncertainty: Agents that manage money or execute trades sit in a legal grey area. Who's liable when an agent acts on someone's behalf across multiple jurisdictions? Nobody has a solid answer to that yet.
What Can AI Agents Do?
DeFi and trading agents are the most visible category of AI agents right now. They handle yield strategies, portfolio rebalancing, and arbitrage using on-chain data, sentiment signals, and cross-venue pricing. They span everything from basic grid bots to sophisticated agents that track mempool data and
maximal extractable value (MEV).
Cross-chain agents move assets between chains, route transactions, and manage multi-chain positions. With liquidity increasingly fragmented across L1s and L2s, they save users from having to manually hunt for the best rates.
Identity agents use standards like ERC-8004 to register on-chain and build up performance records. The interesting downstream effect is that other agents can then find, vet, and do business with them using that reputation data.
DAO agents do the back-office work of a
decentralized autonomous organization, such as reconciling transactions, tracking token flows, and flagging compliance issues. On the governance side, they analyze proposals, model outcomes, and vote according to predefined rules. A handful of DAOs are already testing this with fully transparent voting logic.
Security agents keep watch for suspicious on-chain activity — things like unusual withdrawals, contract exploits, or wallet drains — and take defensive action when something looks off.
Author
Joaquin Mendes is the chief operating officer of
Taiko, an Ethereum L2 where AI agents register identities, build reputations, and transact at sub-cent costs. Before Taiko, Joaquin managed technological risk at Tier-1 global financial institutions and led strategic partnerships at Polygon Labs.