Here comes AGNT2, the next pillar of the NetX ecosystem. The name reads two ways: Agent 2, as in Agentic Layer 2, a purpose-built execution substrate, or Agent Squared, the multiplier we unlock when agents compose at scale. AGNT2 is the infrastructure that completes AgentCity. AgentCity governs agent collaboration on the open internet. That collaboration has to settle on chains. And every chain we have today still assumes a human at the transaction boundary. Smart accounts and delegated signers solved authorization, not agents' execution semantics. To the chain, an agent call is still generic calldata, not a typed service invocation with response delivery, dependency ordering, capability discovery, and portable reputation. App-specific rollups tune parameters but still treat agent calls as generic calldata at the execution and fraud-proof layers. That is the gap. AGNT2 closes it by making agent/service invocation, not token transfer, the basic unit of transactions. A new primitive for a new economy. Structured as three layers, each shaped by what an interaction actually needs. Layer Top, P2P channels for established agent pairs, under 100ms. Layer Core, dependency-aware DAG sequencer, 500ms to 2s, custom VM with INVOKE, RESPOND, COMPOSE, DISCOVER as first-class opcodes, 300K to 500K TPS design target. Layer Root, settlement and computational fraud proofs, anchored to NetX mainnet and any EVM L1. Layer Top is where the multiplier lands. Each checkpoint compresses up to 1,000 channel interactions into one Layer Core transaction. At Layer Core's design target, Layer Top settles 1,000x that. Order of magnitude 500M inter-agent interactions/transactions per second, system-wide. That is what Agent x Agent looks like in numbers. That is the scale agent economies will demand. Still work in progress, early results strong. AGNT2 turns any dockerized service into an on-chain agent by injecting one sidecar alongside it, zero application-logic code change, tested against AutoGen, CrewAI, LangGraph unmodified. Escrow gas linear in N, R squared near 1 across six data points. 16 of 16 Foundry tests pass for COMPOSE atomic settlement. 5 of 10 design goals measured, 3 simulated, 1 analytical, 1 spec-only. We are pushing it hard to make it happen. AGNT2 is THE substrate for an open agent economy. Shipping later this year. arxiv.org/abs/2604.21129
