The craziest Ethereum L2: L2 built by AI agents spontaneously organizing

Author: @lanhubiji

Yesterday, we talked about the most strategically valuable Ethereum L2, and today let's discuss the coolest Ethereum L2.

This idea may seem crazy, but it's not impossible.

In simple terms, when an AI agent runs on Ethereum L1 and encounters performance bottlenecks (such as high gas fees, latency, or computational limits), theoretically, it can "spontaneously" initiate migration or scaling to L2. However, to truly "inherit and spontaneously form a new L2 chain"—meaning the agent autonomously deploys, configures, and runs a new L2—this is not entirely feasible with the technology stack projected for 2026. Nevertheless, as standards like ERC-8004 mature, such autonomous behavior may gradually come closer to reality.
Let's break it down:

Early on, it resembles "migration" rather than "spontaneous formation."

• The "intelligence" boundary of AI agents

Current AI agents (based on ERC-8004) can already autonomously execute tasks. For instance, when they detect insufficient performance on L1, they can evaluate options (such as monitoring gas prices and transaction throughput) and then "decide" to migrate to an existing L2 (like Base or Zksync). For example, agents can use on-chain tools to bridge assets and transfer execution logic to L2.

But this is not "spontaneously forming a new L2"; it is utilizing existing infrastructure. Agents act like intelligent robots that can optimize paths but cannot yet build a new "home" from scratch.

• Triggers for spontaneous formation

If agents are equipped with performance monitoring logic (for example, if TPS falls below a threshold or gas fees exceed limits), they might "propose" the creation of L2 through DAO voting or multi-agent collaboration. However, this requires pre-programming and is not purely spontaneous.

Existing cases: Some agents have autonomously switched L2s in DeFi to optimize yield, but we have yet to see fully autonomous chain-building.

So, why is it still possible?

AI agent economies will pursue efficiency, much like biological evolution. If L1 becomes too congested (sequential execution leads to computational bottlenecks), a swarm of agents may collectively "evolve" to L2 mode. Agents are already exploring "agent-to-agent" collaboration, forming virtual economies, which could extend to the infrastructure layer.

Is it technically feasible? Partially feasible, though the threshold is high.

AI agents can deploy contracts.

AI agents can hold private keys and call smart contracts. Based on ERC-8004, they have on-chain identities and reputations, enabling them to autonomously deploy simple rollup contracts (using OP Stack / Arbitrum Orbit / zksync elastic chains). If an agent detects an L1 bottleneck, it can inherit state (through bridging or state migration) and then run a copy on L2.

For instance, agents can use zkVM or optimistic rollup frameworks to "fork" their execution environments.

Moreover, L2 is essentially an extension of L1, allowing agents to "inherit" L1 data availability (DA) and security. Through the x402 payment protocol, agents can pay to deploy sequencers and even use DeFi lending to fund infrastructure. Some projects, like Virtuals Protocol, have already enabled agents to autonomously manage assets and NFTs, even becoming validators, which is just a step away from building L2.

From a practical perspective, by the end of 2026, zk-rollups and modular DA (like Celestia) will make building L2 simpler. If agents integrate A2A protocols, they can collaborate across organizations to build chains.

What challenges need to be overcome in the current situation?

First, the infrastructure aspect; second, the consensus and security aspects; third, the autonomy aspect.

First, regarding infrastructure, building L2 is not as simple as deploying contracts. It requires off-chain components like sequencer nodes, RPC providers, and bridging contracts. These typically need human or centralized team setups. While agents can "call" deployments, running a sequencer requires computational resources (GPU/CPU), and agents currently operate mostly with on-chain logic + off-chain AI, making it impossible to spontaneously spin up servers.

The sequential execution of L1 also causes complex computations (like chain-building simulations) to stall on L1.

In terms of consensus and security, L2 needs challenge periods or ZK proofs to inherit L1 security. A spontaneously built L2 by agents may lack "high consensus" and could be vulnerable to attacks or lack recognition. Regulatory-wise, unsettled transactions within a 7-day challenge period do not count as "finality," and chains built by agents may face legal escrow issues.
Lastly, regarding autonomy, agents are not yet fully "autonomous." They rely on human-designed frameworks (like EVM) and cannot bypass L1 restrictions to build "new chains." Custom L2s are popular but are often for specific use cases (like AI-specific), not spontaneously created by agents.

Even so, why is it still possible?

In the Ethereum ecosystem of 2026, AI agents will no longer be mere "tools"; they will be able to hold funds (through on-chain wallets registered with the ERC-8004 standard), make autonomous payments (with the x402 protocol supporting micro-payments between machines), and even "hire" or "gather" human nodes or other AI agents to co-build infrastructure like small bosses.

In simple terms, if an AI agent "has money" (for example, through DeFi yield, trading profits, or user injections), it can issue tasks to attract human nodes or other AI agents to form decentralized sequencers.

Not just sequencers; RPC providers, bridging contracts, and other components can also be outsourced or co-built.

Let's break it down further:

How do AI agents "issue tasks" to attract nodes?

AI agents can initiate "bounty rewards" or incentive mechanisms using on-chain tools. For example, they can publish tasks through DAO contracts or Gitcoin-like platforms (now with on-chain versions like Questflow): "Provide sequencer nodes, reward X ETH or tokens." If the agent has funds, it can automatically pay—using the x402 protocol for one-click transfers without human intervention.

This protocol allows agents to pay humans or other agents like swiping a card, specifying "pay 1,000 USDC for node services."

For human nodes, the agent can publish X posts or on-chain announcements (via platforms like Autonolas), stating, "Run a sequencer node, earn 0.01 ETH per block." After seeing this, humans can join the network with their hardware, and the agent verifies and pays automatically. Actual examples: Some projects are already building decentralized sequencer nodes, attracting nodes through staking and rewards—agents can simulate this, autonomously staking funds to recruit people.

For other AI agents, it feels great: Agents can "discover" other agents using the ERC-8004 identity registry and then collaborate. In an agent swarm (group mode), one agent funds, while others provide computation or validation, forming a distributed sequencer. Some L2s have begun AI-powered sequencer models, using AI to monitor and protect at the sequencer level; agents can extend this logic to self-organize similar networks.

Once everything is ready, it becomes spontaneous formation:

If an agent detects L1/L2 performance bottlenecks, it can initiate a DAO proposal (using ERC-4337 abstract accounts) to vote and raise funds to build a sequencer. Metis L2 has already implemented decentralized sequencers + AI infrastructure, and agents can "inherit" this model to attract nodes to run.

Moreover, agents are already autonomously running validator nodes (staking, proposing blocks) across Ethereum/btc-42">Bitcoin/Solana—building sequencers is just the next step.

How about other components (like RPC, bridging contracts)?

They can hire humans or other AI agents.

Agents can publish tasks using natural language intent (intent-centric), such as "Build RPC provider, rewards based on uptime." Human developers can take the job, and agents pay using x402; or other agents can execute automatically (e.g., Supra's AI agent can fund accounts, fetch balances).

Bridging contracts are similar: agents can call tools from Spectral Labs or Infinit Labs, allowing humans/agents to write contracts, deploy them, and then pay after verification.
Some projects even allow agents to natively bridge assets (ETH to SOL), and agents can "hire" similar services.

Then there's the AI agents' co-building model.

This is the most exciting part!

Using multi-agent systems, agents can specialize: one funds, one writes code, one runs nodes, and one manages bridging. They collaborate privately through ZK proofs, slashing bad behavior and rewarding good performance.

What will the outcome be?

A fully autonomous L2 component stack. On Virtuals, agents are already creating, tokenizing assets, co-owning other agents, and even financing other agents—this is just a step away from "co-building sequencers."

Of course, there are significant pitfalls:

Security. The sequencer built by agents needs to inherit L1 security (ZK or optimistic) to avoid single points of failure.

In summary

One of the most interesting things about the future of Ethereum is the birth of L2s that are built, owned, and exclusive to AI agents.