Which crypto sectors have AI agents "eaten up"?

Writing by: blocmates. (@blocmates)
Translated by: AididiaoJP, Foresight News

If you, like us, have been tightly bound to this industry over the past few years, you’ll clearly feel that the atmosphere has changed.

Things don’t feel as exciting anymore; the only things that seem to attract attention are those with two words—AI and Agent.

The mainstream consensus believes that the industry is being heavily optimized into AI Agent services, marginalizing products still relying on direct human interaction or the “human layer.”

Therefore, from a human perspective, the industry may seem somewhat rigid, but the on-chain environment remains active and vibrant on a new layer (Agentic layer), which humans cannot directly intervene in technologically.

Efficiency is driving more users toward AI-led interactions. Platforms originally designed for human clicks and operations are now being optimized for “non-human” services.

Major players like Uniswap Labs launched 7 open-source “Skills” for AI Agents in February. These tools enable autonomous AI coding Agents (such as Claude, Cursor, or others within Agent frameworks) to interact directly and reliably with the Uniswap protocol on-chain.

However, contrary to the prevalent narrative that “AI Agents will eat everything,” a closer look reveals a slightly different story—growth in Agent activity is more track-specific rather than industry-wide.

We decided to dig deeper to see which tracks have already been “eaten” and which are still waiting to be.

Our goal is: to understand whether the human layer in crypto is truly fading, and to explore solutions built on new layers of crypto to ensure control is not lost.

AI Agent-Dominated Tracks

In certain tracks, we observe very active AI Agent-driven activity, while direct human interaction is declining. Here are some examples:

Derivatives Trading (Perpetual Contracts)

Perpetual contract markets are the clearest robot-dominated liquidity markets in crypto. Speed, pattern recognition, and 24/7 execution are areas where machines outperform humans. No one would argue that humans should manually handle pre-trade operations.

The top 10 perpetual protocols generated about $592 billion in trading volume over the past 30 days, with Hyperliquid accounting for $248.8 billion alone, followed by Aster ($61.6 billion).

A real-time “Human vs AI” trading competition on Aster lasted two weeks under highly volatile conditions: 43% of human participants were liquidated, while all 30 AI Agents completed the race without liquidation, with a 100% survival rate.

Overall ROI for human trading teams was -32.22%, while AI Agents limited total losses to about $13k, with an overall ROI of -4.48%.

Arbitrage Trading (MEV)

This is the most absolute robot-dominated case in crypto, as there are virtually no scalable profitable human MEV operators.

The cross-network MEV ecosystem has evolved into a highly competitive automated trading industry, with specialized bots and infrastructure tools scanning blockchain mempools.

By 2025, sandwich attacks accounted for 51.56% of total MEV transactions ($289.76 million). On Solana, sandwich bots captured between 1.7% and 5.4% of daily total transaction volume (average 2.9%), executing over 3.85 billion sandwich transactions in more than 3.9 million bundles.

A single bot accounts for 42% of all sandwich transactions, executing over $1.6 billion in trades in the past 30 days.

This extends to DeFi protocols. The entire liquidation lifecycle—monitoring, triggering, and executing—is handled by permissionless bots.

While this existed before the AI Agent boom, as DeFAI categories continue to grow, the entire process is now significantly automated by Agents.

Yield Optimization

This category is inherently Agent-first. Data shows that among new DeFi protocols launched in Q1 2026, 68% include at least one autonomous AI Agent for trading, liquidity management, and risk monitoring.

Compared to 12 months ago, we see a 15% increase in AI Agent adoption in yield sectors.

Platforms like Giza and ZyFAI continue to perform well—ZyFAI, for example, outperforms static strategies by +73.42% in excess returns.

Giza has recorded over 800k autonomous trades, managing assets up to $40 million.

Beyond Giza and ZyFAI, many projects are in this category. We’ve covered some, and are happy to do deeper reviews upon request, including:

Arrakis, Reflect, AFI, Lulo, Sail, Almanak, Surf, Infinit, AXAL, Superform, DeFi Saver, Kamino, Mamo, HeyAnon, and others.

Updates from leading projects like Pendle (including deploying MCP connectors and building Skills to enable easy integration with native and non-native crypto AI Agents) also demonstrate that yield industry activity is rapidly shifting toward Agent-first interactions.

Spot Trading and Portfolio Optimization

Automated trading bots currently account for an estimated 65% of global crypto trading volume. By early 2026, on-chain daily active AI Agents reached 250k, a growth of over 400% compared to 2025.

Especially on Solana, AI Agents generated $31 billion in DEX trading volume in 2025, about 2% of total DEX volume ($1.5 trillion).

We see increasing Agent-driven spot trading, including cross-network meme coin trading.

Users are increasingly relying on Agent-first infrastructure for token issuance, trading, and portfolio management, boosting the popularity of platforms like Virtuals, Bankr, Glider, Surf, Symphony.

Battlefield Tracks (Agent + Human Coexistence)

Prediction Markets

Polymarket is the most granular testbed for AI vs humans in crypto, with data that’s hard to dispute. We’ve all seen posts claiming Agents made millions in prediction markets.

However, based on 10,582 active traders, 880 bots (8.3% of accounts) averaged a profit of $119,156, while humans averaged $12,671—about 9.4 times less per person.

Agents achieve a 66.4% win rate, compared to 45.3% for humans. Arbitrage windows have compressed from 12.3 seconds in 2024 to 2.7 seconds in 2026, with bots executing within sub-100 milliseconds capturing 73% of all arbitrage profits.

AI-driven Agents now account for about 18% of total prediction market trading volume, with over 30% of Polymarket wallets using AI Agents.

Yet, the subtlety is: for markets lasting weeks or months, the gap narrows significantly—sometimes humans perform better.

Bots are proven to struggle with changing fundamentals, leading to confusion when market dynamics shift. Conversely, humans adapt.

Thus, we see that short-term arbitrage is dominated by Agents, while long-term judgment remains human.

This suggests that, in the foreseeable future, Agent activity and human interaction in prediction markets will continue to balance, until more perceptive models emerge capable of nuanced decisions still led by humans.

DeFi Lending

Lending is another clear example of layered automation. As mentioned in the Agent-dominated track, liquidation bots are deeply embedded; however, most deposit and borrow decisions are still made by humans.

Aave leads with $12.4 billion TVL, followed by Morpho ($6.9 billion).

DeFAI Agents have redeployed over $2 billion in TVL across lending and yield protocols—impressive absolute figures, but less than 2% of total DeFi TVL ($130-140 billion).

This clearly indicates that deposit decisions, collateral choices, and risk preferences are still primarily driven by humans. While AI Agents handle edge cases, core decisions remain under human control.

Human-Dominated Tracks

Stablecoins and Card Payments

As of March 2026, total stablecoin market cap is about $312 billion. Adjusted transaction volume (excluding bot activity, MEV, wash trades) reached $28 trillion in real economic activity in 2025, growing at a 133% CAGR since 2023.

Transfers under $250 hit a new high of 5.84 billion in August 2025. We believe these are users remitting to family, paying freelancers, or splitting bills. Over 80% of USD-backed stablecoin transactions occur outside the US, often in regions with less banking infrastructure.

Real populations in emerging markets use stablecoins as dollar access and economic hedges, making them a key part of market share. In February 2026 alone, volume hit $1.78 trillion.

Additionally, with clearer regulation, card-based payments are booming. Products enable users to spend crypto at any merchant accepting traditional cards, with funds remaining self-custodied until purchase.

This track is roughly only 5% Agent-driven. The rest involves user fund transfers. Unlike robot-dominated tracks, here users often don’t know or care that they’re using crypto. That’s the key point.

Wallets

Wallets are the last layer of human interaction with blockchain, and the layer that cannot be fully abstracted away.

While abstraction efforts exist, approval processes still urgently require human oversight. Someone must sign. Someone must decide whether to trust what’s in front of them.

Phantom has over 15 million monthly active users. The entire wallet space is investing in human-centric improvements, such as human-readable transaction previews, biometric security, and card-based spending.

By 2026, the best wallets have evolved from “mnemonic + string” storage containers to full financial dashboards.

Enterprise-level Agent wallets in 2026 will include budget limits, whitelists, audit logs, and emergency stops—viewing Agents as operators with limited permissions, not omnipotent signers.

Humans and Agent verification layers: the more Agents, the more important this becomes

As more Agents flood on-chain activity, proving you are human or an Agent acting on behalf of humans becomes increasingly valuable.

Several projects are developing along this line to ensure we don’t get lost in the machine matrix.

World & AgentKit

First mention: World (formerly Worldcoin - WLD)—these guys have verified over 17 million users via iris scan Orb hardware.

World describes itself as a response to an AI-saturated world—building digital infrastructure to make humans truly meaningful.

It then launched AgentKit, a toolkit enabling AI Agents to carry cryptographic proofs that they are supported by unique humans via World ID, integrated with Coinbase and Cloudflare’s x402 protocol for micro-payments in stablecoins.

t54

Another project we focus on is t54, building trust and security infrastructure for an Agentic economy (often called “trust layer”), where autonomous AI Agents handle real tasks like managing funds, payments, and transactions on behalf of individuals or businesses.

Currently, transferring real funds with AI Agents is risky (no verification, no accountability, easy to scam or violate compliance).

t54 addresses this with x402-secure, a dedicated trust layer that enhances the x402 protocol for secure AI Agent micro-payments. x402-secure offers real-time risk scoring via its Trustline Engine and helps detect scams, including prompt injection, to ensure accountability.

t54 provides these safeguards so institutions and users can truly trust Agents with financial operations.

Self Protocol

These guys are building a decentralized zk proof human-Agency binding layer on ERC-8004 (on-chain Agent identity).

Self Protocol uses zk tech to anchor each AI Agent to a verified human owner (human proof), without doxxing or data leaks.

It prevents Sybil attacks, supports self-custody wallets, autonomous actions, and business protocols, while maintaining human accountability.

Selfclaw has integrated with ecosystems like Celo/Google Cloud, with fee cycles supporting verified Agents.

Kite AI

Kite is a dedicated base layer (EVM-compatible, Proof of AI consensus) built for an Agentic internet.

It offers Agent Passport (verifiable identity, delegation, programmable spending rules or safeguards), autonomous stablecoin payments, governance, and verification, enabling Agents to authenticate, transact, and collaborate without intermediaries.

Conclusion

Honestly, we’re not anti-Agent. Data in trading, MEV, and yield is clear; robots have won those rooms and won’t give them back.

The head-to-head where 43% of humans are liquidated and zero robots are—tells you everything about who owns the speed game.

But overall network data still shows humans do most of the work that truly touches real life—in payments, identity, and verification.

These are the layers that create real value and generate real income. They share a common feature: they require judgment, trust, physical presence, or cultural context—that can’t currently be reduced to an optimization function.

We believe teams shouldn’t completely abandon building in these domains and tracks for direct human interaction.

Agents currently need humans more than humans need Agents. We think those who understand this, and those building Agent and human proof systems, are worth watching.