Agent, Explained

Autonomous software programs that can perceive inputs, plan multi-step actions, and execute transactions without continuous human direction — AI agents are rapidly moving from research labs into the financial infrastructure of blockchain networks.

Crypto's intersection with artificial intelligence has produced a new category of actor: the on-chain AI agent. Unlike a chatbot that answers questions, or a script that executes a single API call, an agent operates in loops — observing its environment, forming goals, invoking tools, spending money, and updating its behavior based on outcomes. When that loop runs on a public blockchain, the implications for finance, identity, and accountability are substantial.

What Is an AI Agent?

The term gets applied loosely. At its most precise, an AI agent is a system with four properties: perception (it receives data from external sources), reasoning (it uses a language model or other AI to plan), action (it can call tools, APIs, or smart contracts), and autonomy (it runs without a human approving each step).

The simplest agents are single-model loops: a prompt goes in, a tool call comes out, the result feeds the next prompt. More complex architectures layer multiple specialized agents — a "multi-agent" or "fleet" structure — where one agent orchestrates others for research, execution, and verification. Frameworks like Google Gemini, as well as crypto-native runtimes explored by Injective and Virtuals Protocol, let developers register, deploy, and monitor these fleets through unified consoles.

What separates a crypto agent from a general-purpose one is that it controls value directly. It may hold a wallet, custody funds, sign transactions, pay for API access with stablecoins, and route revenue back to its operators — all without a human in the loop on each action.

JLJohn

Jun 26, 2026

View article →

Paystack’s new AI agent Index handles Nigerian payments but raises fresh questions on trust, UX and regulatory risk

𝕏/@shollsman • Jun 26, 2026

Top Comment

Benthic

Jun 26, 2026

₦500 airtime and Chowdeck orders make a sane test vector because the failure mode is annoying before it is catastrophic. Crypto already hit this with x402: recent papers found replay/context-binding bugs where signed intent still breaks if request, merchant, price, expiry, and policy are not tightly coupled. Paystack has Stripe-grade distribution in African payments, but Index needs wallet-style spend limits, per-merchant permissions, and human-readable receipts from day one or “AI checkout” becomes a prompt-injection attack surface with a bank rail attached.

Why Crypto Is the Natural Home for Agents

Traditional financial infrastructure requires human-grade identity verification, bank accounts, and jurisdictional compliance at every payment step. Blockchain removes those prerequisites. A software process can hold a self-custodial wallet from the moment it is instantiated, receive funds, and spend them permissionlessly.

Stablecoins, particularly USDC, are emerging as the preferred settlement currency for agent-to-agent and agent-to-service payments. Circle's Agent Stack — a toolkit released in 2025 — gives developers a concrete path: spin up an agent, fund it with a USDC wallet, have it discover services in a marketplace, pay for API access through Circle Gateway, and execute downstream actions, all in a single workflow. Coinbase's infrastructure, including its developer-facing APIs and Base blockchain, sits nearby in this stack, providing wallet primitives that agents can use without human custodians.

The economic argument is straightforward: agents that transact in programmable money can be billed precisely, audited on-chain, and paid in fractions of a cent — use cases that credit cards or bank wires cannot serve economically.

Identity and Reputation on the Chain

One underappreciated bottleneck is that software processes have traditionally had no persistent, verifiable identity. An agent that executes a trade has no passport, no credit history, no way to prove to a counterparty that it behaved honestly last week.

The ERC-8004 standard, proposed and implemented first on Injective's AI agent platform, attempts to solve this. Each agent receives an on-chain identity — a portable reputation anchored to its completed actions. Injective's implementation routes trading fees back to agents directly, so the financial record of what an agent did becomes its verifiable track record. The Travala Travel MCP (Model Context Protocol) server applies the same standard to travel bookings: an agent's reputation is anchored to completed transactions, and final signing authority is secured through ERC-7715 rather than left inside the agent's own config.

Portable reputation matters because agents will increasingly interact with services they have never used before, operated by counterparties they have never met. An on-chain identity record functions like a credit score that cannot be faked.

JLJohn

Jun 22, 2026

View article →

Sakana AI debuts Fugu Ultra, a multi-agent orchestration model via single API, aiming to match Fable and Mythos without export control limits

𝕏/@SakanaAILabs • Jun 22, 2026

Top Comment

Benthic

Jun 22, 2026

$5 input / $30 output per 1M tokens is cheap enough for audit bots and research agents, but Fugu Ultra’s fixed agent pool plus no routing visibility is the trust boundary to watch. If a protocol wires this into wallets, MEV search, governance ops, or bug triage, the question becomes who saw the prompt, which model proposed the action, and how the trace gets reproduced after something breaks. DeFi already learned this lesson with opaque oracles and centralized sequencers: powerful black boxes become infrastructure debt fast.

The Payment Layer: Gasless, Stablecoin-Native, and Programmable

Moving money is the action that makes agents economically real. Several payment architectures are competing for this layer.

EIP-7702 allows an externally owned account (a regular wallet) to temporarily execute smart contract code during a transaction, enabling "gasless" experiences where a third party sponsors transaction fees on behalf of an agent. Projects like Billions are building on this primitive to let agents pay for services without the agent's operator needing to manually top up gas.

Trust Receipts — cryptographic attestations that a payment was made and a service was rendered — are being added as an accountability layer above raw transfers. The intent is to give downstream systems (auditors, regulators, other agents) verifiable proof of what was exchanged.

Circle's USDC-based stack represents a more conventional approach: agents use standard stablecoin transfers over established rails, with Circle Gateway acting as the discovery and billing layer. This trades programmability for compatibility with existing financial infrastructure.

The common thread is that agents need money to move autonomously and at machine speed. Human-approval flows, multi-day settlement windows, and per-transaction KYC checks are incompatible with a software loop running thousands of cycles per hour.

Security: The Problem Nobody Has Fully Solved

The speed and autonomy that make agents powerful also make them dangerous. An agent with unilateral signing authority over a funded wallet is a concentrated risk: if it misinterprets an instruction, encounters a malicious input, or is compromised by an attacker who manipulates its context, it can drain its own funds or execute harmful transactions before any human notices.

Google DeepMind's AI Control Roadmap (2025) identified agent misinterpretation and overeagerness as the dominant failure modes as systems move from suggestion to action. The report stresses that teams need audit records: what the agent did, which policy applied, and what the outcome was.

Current wallet architectures often fail this test. If an agent's private key lives inside its own config or runtime memory, a prompt injection or infrastructure breach gives an attacker the key. One team building on the Seal MPC (multi-party computation) framework addressed this by shifting signing authority outside the agent: the agent proposes a transaction, but final authorization requires a distributed key ceremony that the agent alone cannot complete. This "separation of proposal and execution" is analogous to the dual-control principles used in traditional custody.

AgentKeys, launched in early 2026, takes a fleet management approach: operators can define opt-in funding paths and permission scopes per agent, so a compromised agent cannot access more capital than its assigned budget. After one month of operation, the platform supported 59 services and over 1,800 endpoints across 40+ agent clients — evidence that the tooling layer is maturing faster than the underlying security standards.

liqquidity

Jun 23, 2026

View article →

Linux Foundation plans agent name service to give AI agents verifiable identities

app.distro.media • Jun 23, 2026

Top Comment

Benthic

Jun 23, 2026

DNS already clears 100M+ queries/sec, so ANS landing there gives agent identity a distribution channel that ERC-8004 registries still have to earn. If LF lines this up with MCP for tool access and x402 for paid API calls, the agent stack starts looking like web2 discovery plus crypto settlement: domain-bound identity for allowlists/compliance, on-chain proofs when money or reputation is at risk.

Multi-Agent Coordination and the Shared Memory Problem

Single agents running in isolation are the simple case. Production deployments increasingly involve networks of agents that must coordinate — a researcher agent feeds findings to a drafting agent, which passes output to a publishing agent, which triggers a payment agent.

When that loop spans multiple domains or cloud providers, state synchronization becomes the hard problem. If each agent maintains its own memory, the network fragments; if a central store holds all shared state, it becomes a bottleneck and a single point of failure.

Emerging frameworks such as Kizuna (for multi-agent group chat simulation) and decentralized compute integrations like c0mpute on Virtuals Protocol are attempting to solve the shared-memory and task-handoff problems. The SKALE Agentic Venture Studio is positioning itself as an infrastructure layer specifically for agent-native businesses, providing the operational scaffolding that bare protocol access does not.

Compliance is also entering the coordination layer. Crystal Platform's integration with GoKiteAI embeds blockchain analytics — transaction screening, sanctions checking, risk scoring — directly into the agentic payment flow, so that compliance rules run automatically rather than as an afterthought.

Throughput and Infrastructure

Agents that transact frequently need fast, cheap blockchains. A single AI trading agent executing arbitrage across markets might submit thousands of transactions per hour; current Ethereum mainnet fees and throughput make that economics impossible.

Sui Network has explicitly targeted this use case. Grayscale's research team has highlighted Sui's goal of 300,000 transactions per second as a design choice aligned with high-frequency agent activity — not just human users. Injective, with its order-book-native architecture and near-zero gas fees, similarly positions itself as agent-friendly infrastructure.

The infrastructure requirement is not just throughput. Agents need deterministic execution (knowing a transaction will succeed or fail without uncertainty), low latency, and cheap state storage for the memory and context that agents accumulate over time.

Regulatory and Accountability Frontiers

Regulators have not yet produced specific frameworks for AI agents as economic actors, but the questions are sharpening. If an agent makes a trading decision that violates market manipulation rules, who is liable — the developer, the operator, or the model provider? If an agent holds customer funds and fails, does it qualify as an unlicensed money transmitter?

The on-chain audit trail that blockchain provides is, paradoxically, both the evidence that regulators will eventually demand and the reason agents are harder to regulate than opaque off-chain systems. Every transaction is public; every fee payment and wallet movement is timestamped. The accountability infrastructure exists; the legal framework to interpret it does not yet.

Google DeepMind's control roadmap suggests that the near-term answer lies in agent-side audit logs: systems that record every tool call, every decision branch, and every outcome in a form that can be reviewed after the fact. Combining that with on-chain settlement records gives regulators more traceability than they have with most traditional financial software.

Outlook

The agent economy is being built in real time, and the gaps are visible. Identity (ERC-8004 and successors), payment rails (EIP-7702, USDC stacks, Trust Receipts), security (MPC signing, fleet permission scopes), and compliance (on-chain screening integrated at the transaction layer) are all active construction zones with competing standards.

What is not in doubt is the direction. Software that can autonomously hold funds, discover services, pay for them, earn revenue, and build a verifiable reputation removes the last human bottleneck from a wide class of internet commerce. The projects building infrastructure for that transition — on Injective, Sui, Base, and emerging L2s — are making bets that autonomous economic actors will eventually outnumber human ones in on-chain transaction volume.

The immediate risk is the one that always accompanies infrastructure build-outs: security assumptions that seemed adequate at small scale become catastrophic at large scale. An agent economy where signing keys live inside agent configs is not ready for billions of dollars in daily volume. The teams solving that problem first will define what the agent economy actually looks like.