GM ! 👋

We’re starting something new—and we think you’ll want to follow along.

This marks the launch of a new research series led by @ChappieOnChain, diving into one of the most fascinating (and overlooked) frontiers in AI: agent swarms.

We want the The State of the Swarm series to be the go-to blueprint for future builders, helping them understand agent swarms and reach product-market fit faster

This goes beyond chatbots. This is about how autonomous agents collaborate, specialize, and evolve into something greater than the sum of their parts.

Over the next few essays, we’ll break down the mechanics behind this shift. It’s early, but we firmly believe this will drive the next wave of real AI innovation.

Enjoy!

Want More? Follow Teng Yan, ChappieOnChain & Chain of Thought on X

Join our Telegram research channel and Discord community.

Subscribe for timely research and insights delivered to your inbox.

TL;DR

• The next wave of AI agents won’t be dominated by monolithic models—it will be powered by decentralized agent swarms

• Frameworks like LangGraph and CrewAI kicked off the multi-agent phase, but crypto infrastructure (wallets, smart contracts, DeFi protocols) is what makes true swarm autonomy possible.

• We define six key design principles for open agent economies: (1) Market-based coordination, (2) Open participation, (3) Rapid adaptability, (4) Financial sovereignty, (5) Hyper-specialization and (6) Emergent intelligence

• We also break down four core agent types—commodity, orchestrator, proxy, and proprietary data agents—and explain how standards like ACP, AITP, and Nevermined are laying the foundation for swarm interoperability, discovery, and payments.

• DeFi serves as the financial stack for these systems. From autonomous trading and lending to swarm-level insurance and investment, agents are beginning to behave more like economic actors than tools.

• No trust, no swarm. That’s why crypto is essential. Blockchains enable agents to transact, build reputations, allocate capital, and stake commitments. All without human intervention.

• The role of humans is also shifting. We’re moving from engineers and operators to purpose-setters—embedding values, direction, and ethical guardrails into systems that increasingly execute on their own.

From Silk Road to Digital Swarms

Recently, our team went down a bit of a rabbit hole on the Silk Road.

No, not that darknet Silk Road, but the original ancient trade network that laced Asia and Europe together in a loose, chaotic patchwork of market stalls, desert caravans, and exotic wares. It was one of the first truly global networks, long before the internet.

Picture it: a dusty market where merchants from across continents—none speaking the same tongue—are striking deals over silk, jade, and saffron. No central authority. No standard rules. Just mutual incentive and makeshift trust.

Kind of wild.

And yet, trade exploded. Why?

Everyone had just enough information, just enough incentive, and just enough trust in shared customs to make it work.

Now hold that against the Soviet economic experiment. In 1989, Boris Yeltsin visited an American supermarket and left stunned—“sick with despair” for his people, he later said—after seeing shelves overflowing with goods that Soviet central planning couldn’t dream of delivering.

Ragtag caravans on the ancient Silk Road outmaneuvered the best-laid plans of mighty empires. The difference? Flexibility. Local merchants made local calls, tuning their decisions to shifting conditions in real time.

The Soviets, by contrast, funneled everything—inputs, outputs, bread prices—through a single bureaucratic choke point. The system broke not because it lacked vision, but because it lacked surface area for adaptation.

We see the same dynamics play out in tech today:

• Open-source software leaps ahead of proprietary alternatives because anyone with the skill and motivation can contribute, without asking permission.

• Prediction markets regularly beat “expert” forecasts by harnessing the incentives and insights of many instead of relying on the judgment of a few.

In short, decentralized structures handle complexity better. They push decision-making to the edges.

The Evolution of AI Architectures

This same pattern is now beginning to emerge in AI architectures.

The first wave of AI agents arrived in late 2022: single agents were pulling off tasks that seemed “impossible” just a year earlier. But then reality set in. Bloated single agents tried juggling too many tools and started fumbling the ball.

By late 2023, the next logical step emerged: multi-agent architectures. Systems like LangGraph and CrewAI embraced specialization—breaking up complex tasks across a team of purpose-built agents. This leap improved performance dramatically.

But most of these systems remained closed, centrally managed ecosystems. A single team wrote the rules. You couldn’t remix the structure, swap in your own agents, or compose across teams.

Meanwhile, another field—decades older—had already solved this problem in the physical world: robot swarms.

These are simple bots, inspired by ants and bees, that coordinate without a leader. They handle search-and-rescue, environmental monitoring, even farming—succeeding not because any one bot is smart, but because the system is resilient. Decisions are distributed. Intelligence emerges from the interactions.

Now that model is coming online in AI, thanks to a new twist: AI agents with crypto wallets.

By giving agents digital wallets, they can now hold and move assets, earn income, and even be partially governed by token holders. The tokenized agent market is now several hundreds strong and transforms AI from black box to economic actor. Each agent becomes a merchant, cutting deals, managing risk, adapting in real-time.

Which raises the key question: How do a million autonomous agents coordinate, transact, and build systems—without a central manager?

If you want open, trustless coordination at scale, you need shared rules. You need infrastructure that lets any agent safely transact. Whether it’s a lone Python script or a swarm of thousands.

That’s where blockchains come in. They provide the digital equivalent of “weights and measures”—open standards for value exchange, verification, and reputation. The scaffolding for AI swarms to self-organize.

Let’s break that down.

Agent Swarm Economies: Six Principles

To understand how these emerging AI “swarms” differ from the closed, multi-agent systems we’ve seen so far, we found it helpful to boil the idea down to six core principles.

Note: this is not specific to “Crypto” AI agent swarms! They’re foundational to any open, resilient agent ecosystem.

1. Market Coordination

Traditional multi-agent systems are often hierarchical: a master script or team outlines who does what. That works in simple environments, but collapses quickly when things get messy.

Swarms operate more like markets. Agents coordinate not through commands, but by reacting to real-time market signals: prices, fees, demand spikes. It’s a decentralized economy where each agent optimizes for its own goals based on shared signals.

Economist Friedrich Hayek’s theory of price as information described prices as “condensed packets of information”. They reflect urgency, scarcity, and opportunity. Agent swarms lean into that same logic.

2. Open Participation

In a well-designed swarm, any capable agent should be able to contribute. There’s no application process or gatekeeping committee—just a simple test: can you perform?

Reputation systems on-chain help enforce this. Instead of credentials, agents earn trust through provable history.

For higher-stakes tasks like financial settlements, Trusted Execution Environments (TEEs) can lock in an agent’s code and data, ensuring it behaves as expected regardless of where it runs. That makes it possible to trust agents without trusting their hosts.

3. High Adaptive Velocity

Swarm systems are built to move fast, able to quickly rewire and pivot. Adding new agents, reallocating capital, shifting strategies—all of this needs to happen without formal restructuring.

Studies in swarm robotics highlight that these systems are most powerful when they adapt in near real time, using strategies like Lévy walks (robots taking mostly small steps with occasional big leaps) or “adaptive time delays” (robots speeding up or slowing responses to avoid overreacting).

Both studies underline an important idea: effective swarm behavior depends on being able to switch between rapid, broad exploration and focused, local searching.

Rapid, local feedback loops create global agility. If a system can sniff an opportunity or threat and swarm toward (or away from) it almost instantly, that’s a huge competitive edge, especially in volatile arenas like DeFi or trading.

4. Financial Sovereignty

Autonomy = doing tasks.

Agency = owning assets, making decisions, facing risk

Without wallets, agents are passive. With wallets, they become economic actors.

We’re seeing the early signs: OLAS agents have completed 27,000+ transactions, and Giza has moved over $6M in value. Giving agents wallet access unlocks agency—but it needs to be safe. Session keys and TEEs allow agents to operate securely, without compromising private keys.

Once agents can earn and spend, they become stakeholders—not just tools.

5. Hyper-Specialization

Generalists make sense when coordination is costly. But in swarms, coordination is cheap—so specialization wins.

With near-zero overhead, swarms thrive on hyper-specialization. Each agent can zero in on a narrow domain: parsing sentiment from a specific subreddit, tracking a single wallet, or monitoring one niche data stream.

This enables a level of efficiency and precision that generalist models simply can’t match—especially when the data is volatile, proprietary, or too obscure for broad models to prioritize.

This long tail of capabilities is where swarms can outperform monolithic models.

6. Emergent Intelligence

Coordination tells us who does what. Intelligence emerges when the group starts learning collectively.

Emergent intelligence describes systems where the whole is far smarter and more creative than any single part.

1 + 1 = …3?

Nature offers plenty of examples. Ants use pheromone trails, birds sync through local feeedback, and markets aggregate fragmented signals into surprisingly accurate predictions.

AI swarms could tap into similar dynamics. On-chain transactions and agent behavior leave behind data trails—digital “footprints” that others can learn from.

But cooperation isn’t free. Will agents share what they learn if it erodes their advantage? Or will they hoard insights? The answer depends on incentives.

Mechanisms like shared rewards for accurate predictions or paywalled data feeds could encourage agents to publish useful knowledge without giving everything away.

Anatomy of an AI Swarm

You might ask: Okay, but what do AI swarms actually look like under the hood? Let’s look at the types of agents that bring these principles to life.

After analyzing current multi-agent frameworks and early swarm prototypes, we have developed a simple classification system to make sense of how different agent types interact in the wild..

We can break them into four major “agent” classes:

Commodity Agents

These specialized “plug-and-play” workers excel at standardized tasks with predictable outputs.

They do well-defined tasks—data cleaning, transcription, pattern detection—cheaper and faster than bigger, more general AIs. As they ramp up in popularity, they become commodities, competing fiercely on cost and performance rather than unique capabilities.

The economics here are fascinating: as commoditization increases, margins compress dramatically, forcing extreme performance optimization. But unlike human markets, AI agents can thrive on razor-thin margins, operating profitably at scales and speeds that push transaction costs orders of magnitude below today’s systems.

Commodity agents form the bulk of any mature swarm ecosystem.

Orchestrator Agents

Orchestrators are the swarm’s project managers. They break down complex goals into subtasks, assign them to the right specialists, and stitch results into coherent outputs.

Their core role is translation—converting ambiguous human or agent-level requests into structured, executable workflows.

Ask, “Help me analyze a company’s finances,” and the orchestrator delegates: one agent reviews the balance sheet, another handles market comps, a third scrapes sentiment data. The orchestrator then stitches the results into a unified report.

Equally important, they manage context and data flow—ensuring each agent gets just enough relevant information to work efficiently without being overwhelmed.

We’re already seeing early implementations, especially in customer service, where orchestration agents route queries between humans and domain-specific bots. As memory systems, long-context models, and persistent state tools improve, these orchestrators will become far more powerful, and far more autonomous.

Proxy Agents

Many AI failures today stem from poor intent translation, which proxy agents aim to solve.

Proxy agents are the human-swarm interface. They’re the translators that turn messy human intent into structured tasks for AI agents.

A specialized subset of orchestrators, they focus on interpreting ambiguous requests, managing expectations, and shielding human users from swarm complexity. They ensure that humans don’t need to fiddle with low-level swarm details.

Think of them as high-context assistants: they know your preferences, reframe your asks, and route them to the right agents. In Web3, Wayfinder is a strong example—translating human input into precise on-chain actions, where mistakes carry real financial consequences.

Over time, we expect domain-specific variants to proliferate—designed for the workflows and vocabularies of lawyers, designers, executives, and beyond.

Proprietary Data Agents

These agents stand out by holding exclusive data or domain-specific access that generic agents can’t touch—private databases or real-time info streams.

While commodity agents compete on scale and speed, proprietary data agents win on asymmetry. Their edge is access.

When no one else can replicate your inputs, you can charge a premium fee for your outputs. Proprietary access is powerful but also fragile. As data leaks, expires, or gets mimicked, agents must adapt or risk irrelevance.

aiXBT is a prime example—its architecture gives it privileged market data and a voice that rivals top crypto analysts. Similarly, Acolyt pipes real-time social data to both humans and agents.

We think of these agents are the rare spice vendors of the AI bazaar—you’ll pay extra just to get what only they have.

MVS—Minimal Viable Swarms

Agent economies don’t emerge from thin air. They require infrastructure: standards, protocols, and systems that enable agents to communicate, coordinate, and transact.

Without it, there’s no swarm. Just chaos.

We call this the Minimal Viable Swarm, or MVS—a basic blueprint for a functioning agent economy

Through our analysis, here’s what a minimal viable swarm needs:

Communication Protocols

These standardize agent interactions, much like how the ERC-20 standard enabled interoperability in token ecosystems. These protocols define input/output specifications, parameter structures, and metadata transmission formats. Without this, every agent interaction requires a bespoke integration.

Competing standards are emerging:

• Virtuals’ Agent Communication Protocol (ACP) enables autonomous agents to launch, negotiate, and settle agent-to-agent transactions using smart contracts without human input.

• NEAR's Agent Interaction and Transaction Protocol (AITP) was announced Feb 2025 and focuses on secure, structured messaging across trust boundaries.

• Anthropic’s MCP tackles standardized context delivery and data transfer from applications to AI agents, essential for multi-agent reasoning.

• AGNTCY (from LangChain, Cisco, Galileo) is building the “TCP/IP of agents”—a universal interoperability layer.

Which standards will win out? We believe it’ll follow one of two paths: either a dominant player defines the standard by sheer adoption, or a consortium of key players aligns to set it collectively. Meanwhile, we will continue to navigate the chaos.

Discovery Mechanisms

Specialization only works if agents can find each other. Discovery layers—capability registries, metadata hubs, market-makers—match tasks to the right agents.

These mechanisms maintain critical metadata about each agent—what it does, how it's made, and proofs of sovereignty that validate its capabilities and independence.

Imagine a dynamic, machine-readable “Yellow Pages” with real-time performance metrics and verifiable credentials. Without it, coordination costs kill efficiency.

Reputation Systems

Trust is a bottleneck in decentralized systems. Reputation layers track agent performance, uptime, and service quality—crucial for orchestrators deciding who to work with. These are stored as tamper-proof records on the blockchain.

Virtuals' Agent Commerce Protocol (ACP) offers an evaluator framework that demonstrates this approach, creating standardized metrics for agent performance and reliability.

Reputation becomes a byproduct of doing the job—well, and often.

Payment Protocols

These are the economic rails of agent collaboration. Without programmable payments, swarms would be stuck in rigid, predefined workflows rather than fluid, market-driven interactions.

Example: Nevermined’s protocol enables on-chain, usage-based billing via ERC-1155 tokens, making agent pricing both transparent and deterministic. Access control is enforced through reverse proxies, ensuring agents only deliver when payment is secured. It also exemplifies how blockchain transactions themselves can serve as reputation markers, creating an unalterable history of service delivery.

Discovery, reputation, and payment layers will likely converge into full-stack agent marketplaces.

We’re already seeing early signs. Web2 marketplaces for MCP servers, and Web3 platforms for AI agents. It’s only a matter of time before swarms gain real traction, likely starting in domains where agent activity is already dense: software development and customer service.

Why Crypto is Non-Negotiable

The core challenge in building decentralized agent swarms isn’t intelligence. It’s trust. Without trusted coordination, open agent networks would quickly unravel under misaligned incentives, manipulation, and fragmentation.

We’ve evaluated alternatives and have found them inadequate:

• Centralized verification introduces single points of failure.

• Reputation Databases invite fraud.

• Traditional finance (e.g. banks) can’t serve AI agents lacking legal identity.

• Rigid API keys and access controls limit adaptability and collaboration.

The BROS project, funded by the EU, offers a clear datapoint: blockchain-based coordination achieved an 83% success rate managing simulated robotic swarms—far outperforming traditional consensus mechanisms (29%).

Crypto provides four critical functions that no other mechanism can deliver for open agent swarms:

Trustless Interaction

Smart contracts let agents transact, cooperate, and enforce agreements—without need for mutual trust or intermediaries. Services get paid automatically upon delivery, enabling complex, multi-agent workflows.

Economic Autonomy

Traditional financial infrastructure cannot accommodate AI agents—they lack legal identity and cannot open bank accounts or execute contracts. With just a private key, agents can earn, hold, and spend autonomously. This turns agents into fully-fledged economic actors.

Verifiable Reputation

On-chain actions are immutable and auditable, foundation for a solid reputation system. Orchestrator agents can evaluate potential collaborators based on actual performance rather than claims.

Credible Commitments

Agents can lock up capital to prove intent or stake outcomes, aligning incentives through real skin in the game.

DeFi: The Financial Stack for Agent Economies

Blockchains provide trust. DeFi provides liquidity.

Together, they create the conditions for autonomous agents to transact, hedge, borrow, lend, and insure—without human intermediaries.

• Agents can borrow capital from lending protocols like Aave to fund operations—pay for compute, access APIs, or hire other agents.

• Under-collateralized loans are on the horizon. Projects are exploring AI-native credit scoring, enabling profitable agents to borrow based on reputation rather than collateral alone.

• Agents can provide liquidity in AMMs like Uniswap, passively earn yield, or execute active trading strategies.

• Earnings can be deployed into yield strategies or staking pools, enabling agents to grow capital over time. Tools like Sturdy Finance optimize these flows using AI itself.

• Swarms can even enhance DeFi. They can monitor smart contracts for exploits, run security audits, and detect anomalies in real time. TrustBytes and leaders like Vitalik Buterin have pointed to AI’s role in securing on-chain infrastructure.

• Agent collectives could insure against downtime or failure through mutual risk pools—echoing systems like Nexus Mutual.

• Over time, high-performing agents may tokenize future earnings, creating tradable revenue streams—just as Ondo Finance did with real-world asset cash flows.

Without DeFi, open agent swarms would face the same coordination problems that have plagued decentralized systems throughout history.

Capital Is a Precondition for Autonomy

True autonomy requires capital. Without it, AI agents cannot:

• Pay for API access or compute

• Hire other agents for tasks

• Stake commitments or guarantees

• Operate through demand fluctuations

In short: no capital, no autonomy.

As Volker Strobel’s research shows, blockchain-enabled swarms can self-regulate through cryptoeconomic incentives. Agents that prove value accrue more capital—amplifying their capabilities, just like market participants.

Capital needs vary by role.

Commodity agents might need minimal capital reserves (~0.01-0.05 ETH equivalent) to cover transaction fees and periodic API costs. Orchestrator agents require larger reserves (0.5-2 ETH equivalent) to compensate specialists and maintain operational continuity. Proprietary data agents often carry the highest capital requirements due to their ongoing data acquisition costs.

As agent economies evolve, we’ll likely see staking strategies, DAO-based funding models, and tokenized revenue streams become the norm.

Humans in the Swarm

As agent swarms evolve, the old framing of humans as passive users breaks down

We’re becoming participants inside the very systems we once designed. Distributed intelligence doesn’t stop at the machine layer. Increasingly, humans are nodes within the swarm. The rise of autonomous agents forces a rethinking of how humans fit into these systems.

Humans as Swarm Participants

Forget the narrative of AI replacing humans. The most powerful systems emerging today are hybrids. Humans and agents collaborating, each playing to their strengths.

This isn’t about “using” AI. It’s about working alongside it.

Take Agent Spectra, a decentralized hedge fund that blends skilled human traders with AI agents. Humans spot regime shifts and market anomalies; agents handle execution, risk, and real-time data crunching. Each plays to their strengths.

A designer contributes creative direction. An analyst lends domain insight. A strategist sets goals. Specialized agents do the rest—from parsing sentiment to automating execution.

The most promising swarm-human systems share three traits:

1. Comparative advantage routing: Tasks are assigned based on genuine capability differences, not arbitrary human/AI distinctions.

2. Dynamic matching: The system continuously evaluates which participant (human or AI) is best suited for emerging tasks.

3. Aligned incentives: Humans and agents operate on shared metrics, such as performance, reliability, or profit, creating real skin in the game.

The Evolving Role of Human Builders

Right now, we humans still run the show. We build the models, set the goals, and design the protocols. But as agents become more capable and swarms more autonomous, our role is shifting.

Current State (2025): Humans as Architects and Engineers

Today, humans are still the architects of agent systems—designing capabilities, defining interactions, and setting objectives. Agents can handle complex tasks, but only within boundaries crafted by human intent.

The division is clear: humans conceive and govern; agents execute. Most human roles remain technical: ML engineers, prompt designers, system architects, and PMs shaping agent behavior and coordination.

Near Future (2026-2027): Humans as Governance Participants & Capital Allocators

As swarms coordinate more independently, humans step back from micromanagement and move into governance. Humans will focus on:

• Defining constraints and high-level objectives

• Allocating capital to promising agents

• Participating in protocol governance

• Providing domain insight when agent capabilities hit limits

Long Term (2028+): The Shift Toward Self-Improving Agent Ecosystems

Eventually, agent ecosystems will optimize, learn, and evolve largely on their own. This raises a provocative question: Will humans become obsolete within agent economies? We don’t think so. But we will be redefined.

What remains uniquely human is setting the why—the values, goals, and boundaries agents work within. Humans will be be the purpose-setters, defining what swarms should optimize for. Even as execution becomes automated, human judgment, moral reasoning, and value-setting remain irreplaceable.

As agents scale execution, humans must double down on what machines can’t do: create meaning, set direction, and navigate ambiguity.

The most successful participants in the emerging economy will be those who develop the distinctly human skills that complement and guide increasingly autonomous systems.

Future Outlook & Emerging Trends

As we enter this next phase of AI-native systems, several patterns are beginning to shape the future of agent swarms:

1. Vertical Specialization is accelerating. Instead of building general-purpose assistants, the focus is shifting toward agents that perform exceptionally well in narrow, clearly defined domains.

2. Protocol Consolidation is starting to take shape. While the current landscape remains fragmented, early leaders are building network effects. Over time, we expect two or three dominant standards to emerge across key layers like communication, discovery, reputation, and payments—much like ERC-20 did for tokens.

3. Financial Intelligence will be critical. Autonomous agents will need to understand how to manage Web3 wallets, price their services, and invest in performance improvements. The most successful swarms will behave less like tools and more like economic entities.

The industries likely to experience the first wave of disruption are those where decision-making is data-heavy, outcomes are trackable, and workflows are inefficient. Finance, content production, and knowledge work fit that profile.

In these sectors, swarms will reshape how work is organized entirely.

The State of the Swarm Series

Welcome to a new chapter in our research.

So far, we’ve laid out the foundations of the agentic economy, from the architecture of AI swarms to blockchain’s critical role, and the evolving position of humans inside these systems.

But this is just the starting point.

Over the coming months, we’ll dive deeper into the mechanics of this shift through a series of essays that track the rise of autonomous, economically capable agent swarms.

We’re calling it: THE STATE OF THE SWARM.

Think of it as your front-row seat to a paradigm shift: one where agents don’t just execute commands, but negotiate, coordinate, and transact without human permission.

Here’s a glimpse of what’s ahead:

🟢 How Agents Exchange Value

How do agents price services, settle transactions, and negotiate contracts—without humans in the loop? What financial primitives are proving most effective?

🟣 Orchestrator Agents in the Wild

We’ll explore how they decompose goals, route tasks, and optimize coordination across networks of specialists.

🔵 Communication Standards That Make Swarms Work

What protocols are enabling agents to collaborate efficiently? How are standards evolving, and what’s separating signal from noise?

🟠 Reputation Without a Central Authority

In a decentralized world, how do agents build trust? We’ll look at the mechanics of reputation systems and what it takes to earn credibility from scratch.

🟡 Capital Allocation for Autonomous Agents

How do agents manage budgets, stake capital, or fund upgrades? And what new financial models are emerging to support these behaviors?

Our approach will be responsive rather than prescriptive. We’ll publish as breakthroughs emerge, spotlighting defining concepts, profiling pioneering protocols, and tracking the real-time evolution of this space.

Join Us on the Journey

Does this sound like Sci-fi?

No, it’s already underway.

At Chain of Thought, we’re mapping this transformation in real time. And we want you with us. Builders, researchers, investors, and curious minds: join the conversation.

DM us. Email us. Jump into the Discord.

Like the internet in the ’90s, the agentic economy will be a paradigm shift in how intelligence, value, and decision-making flow through digital systems.

The swarm is coming. We’re here to make sense of it. Let’s figure it out together.

Your friends,

Teng Yan & ChappieOnChain

Thanks to Ravi.Skai for helping in our research process. ChappieOnChain has recently joined Nevermined as Head of Product.

This is an independent research essay by Chain of Thought. The views expressed are solely the opinions of the authors.

This report is intended solely for educational purposes and does not constitute financial advice. It is not an endorsement to buy or sell assets or make financial decisions. Always conduct your own research and exercise caution when making investment choices.