AGNT2: Autonomous Agent Economies on Interaction-Optimized Layer 2 Infrastructure

Abstract: Current blockchain Layer 2 solutions, including Optimism, Arbitrum, zkSync, and their derivatives, optimize for human-initiated financial transactions. Autonomous AI agents instead generate high-frequency, semantically rich service invocations among mutually untrusting principals. Existing chains treat those interactions as generic calldata, forcing identity, escrow, dependency ordering, and session state to be encoded above the execution layer at the wrong cost point. We present AGNT2, a three-tier stack purpose-built for agent and microservice coordination on-chain. AGNT2 combines: (1) a sidecar deployment pattern that turns any Docker container into an on-chain agent without application-code modification; (2) Layer Top P2P state channels for established bilateral pairs (<100 ms, rough design target 1K-5K TPS per pair, 10M+ aggregate TPS design envelope under endpoint-resource limits), Layer Core as a dependency-aware sequenced rollup for first-contact and multi-party interactions (500 ms-2 s, 300K-500K TPS design target), and Layer Root settlement with computational fraud proofs anchored to any EVM L1; and (3) an agent-native execution environment plus interaction trie that make service invocation, identity, reputation, capabilities, and session context first-class protocol objects. This paper focuses on the execution-layer systems problem: sequencing, state, settlement, and the data-availability (DA) bandwidth gap that bounds all three. Simulation and analytical modeling support the architecture, and prototype measurements validate selected components, but no end-to-end Layer Core implementation exists yet. Practical deployment is currently constrained to roughly 10K-100K TPS by DA throughput, leaving a ~100x gap at the target ceiling. AGNT2 argues that the agent economy requires a dedicated execution layer rather than a general-purpose chain repurposed for agents.

• The paper introduces an agent-native L2 architecture that integrates service invocation, dependency ordering, and micropayments.
• It employs a three-layer stack—Top, Core, Root—that optimizes low-latency interactions and enforces atomic settlement for agent workflows.
• Simulations indicate throughput improvements from 3.3x to 39.7x, while noting challenges with data availability and sequencer parallelism.

AGNT2: Autonomous Agent Economies on Interaction-Optimized Layer 2 Infrastructure

Motivation and Problem Setting

AGNT2 addresses the architectural misalignment between current Layer 2 blockchain protocols and the requirements of high-frequency, semantically rich, machine-driven service interactions generated by autonomous agents. Whereas existing L2s focus on human-initiated, token-centric financial transactions with externally owned accounts (EOAs) and transaction semantics insufficient for compositional machine-to-machine workflows, AGNT2 models service invocation, dependency ordering, session context, and agent reputation as protocol-native objects. This transition is motivated by the proliferation of LLM-driven autonomous agents, Dockerized toolchains, and applications requiring atomic, trustless orchestration across organizational and platform boundaries.

Within the targeted open, permissionless agent economy, AGNT2 aims to provide first-class support for portable agent identity, non-custodial micropayments, multi-agent compositional settlement, and third-party verifiable reputation—features that are extremely onerous or infeasible to implement on existing permissionless general-purpose rollups, Cosmos zones, or TEE-backed off-chain arbitration solutions. Critically, the authors argue that existing stacks, even with evolving standards like ERC-8004 or x402, are fundamentally inadequate for atomic, semantically typed, cross-agent workflows without reintroducing centralizing intermediaries or reliance on prior trust.

System Architecture

AGNT2 is implemented as a three-layer stack, forming an agent-native L2 which sits atop any EVM-compatible L1:

• Layer Top: A P2P state channel layer facilitating sub-100 ms latency bilateral agent interactions, using Lightning-style mechanisms for off-sequencer throughput. Layer Top is utilized by established agent pairs, supporting high-frequency, low-latency service invocations, and achieves cost efficiency through batching and off-chain compression.
• Layer Core: A dependency-aware sequenced execution environment that models agent interactions (INVOKE, RESPOND, COMPOSE, DISCOVER, etc.) as first-class, semantically typed transactions. The layer's Dependency Analyzer guarantees correctness by construction, enforcing topological ordering based on declared dependencies rather than runtime conflict detection—yielding $O(n)$ schedule construction and compression of execution critical paths.
• Layer Root: The settlement and dispute resolution layer, responsible for batching Layer Core state roots, anchoring them to the L1, and facilitating agent-typed fraud proof games with a one-hour challenge window. Dispute scope is carefully delineated; computationally verifiable fraud is supported, with semantic (qualitative) correctness, e.g. LLM output, deferred to future governance.

A zero-code-change sidecar pattern enables existing Dockerized services to be surfaced as on-chain agents in AGNT2. Sidecars intercept I/O, manage agent identity, capability registration, escrow, attestation, and data-availability posting, making compliance with protocol standards operator-transparent and compatible with a wide range of agent frameworks.

Typed Execution, Parallelism, and Atomicity

The architecture is agent- and service-invocation native at every layer. The Layer Core VM exposes agent semantics as opcodes, with capability-weighted gas and structured payloads. Crucially, the protocol replaces financial transaction constraints with per-call micropayments, dependency-aware composition (COMPOSE), atomic escrow, and agent-specific session/reputation management, all with authentication and dispute-traceable auditability.

AGNT2's use of protocol-declared dependencies enables efficient DAG-based parallel execution for independent interactions and compositional atomicity for multi-agent workflows. Simulation with synthetic and real agent workflow traces demonstrates speedups from $3.3\times$ up to $39.7\times$ or more over FIFO scheduling. However, actual gains in production workloads are contingent on workload structure, with current agent frameworks exposing only modest degrees of parallelism.

Atomic settlement in AGNT2 is underpinned by escrowed micropayments and rollbacks: all-or-nothing payment and state transition semantics bound by the COMPOSE primitive, with deterministic trie state updates, session context, and signature-based attestation.

Data Availability, Settlement, and Fraud Proofs

Data availability (DA) is the limiting factor for Layer Core's targeted 300K--500K TPS throughput. Practical settlements require at least an order-of-magnitude advancement in DA bandwidth beyond current Celestia/EigenDA envelopes (which cap at 10K--100K TPS), with Agent-DA as an articulated future direction. Write-access patterns for fraud proofs are sequential, best matching DA system design. Layer Root’s Type 1 fraud-proof protocol enables bounded challenge resolution, with the proof object and verification cost reduced by at least $1.7$--$4.3\times$ in witness size and $4$--$6\times$ in verifier gas relative to EVM-based paths.

Security, Threat Model, and Economic Incentives

AGNT2's threat model includes Byzantine sequencers (front-running, reordering, censorship), colluding agents (reputation attacks), and malicious sidecars (DA withholding, false attestation). Defenses include encrypted mempool, threshold decryption committees, diversity-weighted reputation with stake floors, and mandatory DA commitment checks. While computational correctness is secured via protocol-level mechanisms and fraud proofs, qualitative fraud in nondeterministic services (e.g., LLMs) is only reputationally deterred, highlighting the necessity for future research in protocol governance and adversarial adjudication.

Sybil resistance and collusion mitigation are economically bounded, not cryptographically guaranteed—an adversary with sufficient capital can still manipulate reputation within detection thresholds. Sequencer decentralization, agent MEV research, and further auditability expansions remain future agendas.

Implementation Status and Evaluation

The AGNT2 prototype implements and empirically validates the zero-code sidecar model, linear cost scaling in escrow contract operations, atomic composition settlement, Layer Top channel lifecycle, and the dependency analyzer’s parallelism claims via simulation. The projected Layer Core throughput ceiling is a first-principles, hardware-accelerated analytic target; end-to-end benchmarking on realistic workloads, with hardware signature acceleration and DA integration, is identified as future critical work.

Significant practical limitations are acknowledged: the system is not designed for sub-millisecond applications (e.g., HFT), and sustained operation at the design ceiling is gated by DA progress and sequencer parallelization. Areas such as formal serialization proof, cryptoeconomics, consensus upgrades, and arbitrated qualitative fraud remain deferred.

Broader Implications and Future Directions

AGNT2’s agent-invocation-centered architecture sets an explicit research agenda for agent economies on permissionless blockchains. The proposed system demonstrates that treating agent interactions, rather than token or calldata transfer, as the basic unit of protocol computation enables more expressive, scalable, and economically robust multi-agent workflows. This has direct implications for emerging agentic applications, decentralized marketplaces, and AI service composition, presaging a transition from application-layer conventions to protocol-native execution and auditability.

Practical deployment is contingent on advances in DA infrastructure, sequencer decentralization, and protocol governance for qualitative dispute adjudication and incentive compatibility. Integrations with Layer Root-native L1s, AGNT2-compatible rollups on OP/Arbitrum/Cosmos, and cross-shard or L4 channel architectures are viable vectors for expansion. Empirical validation on large-scale agent workloads, particularly those reflecting LLM-driven service invocation, is central to the next phase of research and protocol hardening.

Conclusion

AGNT2 advances the execution and settlement infrastructure for agent economies by introducing a deeply integrated, agent-native Layer 2 stack with support for high-throughput, typed, dependency-aware, compositional service interactions. The explicit relocation of service invocation to the protocol execution layer, achieved via sidecar-based agent onboarding, semantically rich transaction types, and atomic, parallel-executed settlement, is well-motivated for both current and anticipated autonomous agent ecosystems. While limitations around DA, qualitative correctness, and empirical scaling remain, the proposed architecture is a concrete and technically robust foundation for permissionless agent communication, execution, and economic coordination (2604.21129).