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

Join our Telegram research channel and Discord community.

Subscribe for timely research and insights—delivered to your inbox.

What if crypto’s agent economy wasn’t just a mix of lone-wolf agent traders, but AI agents working together in perfect sync?

In Part 1, we explored how Spectral’s Syntax platform made it ridiculously easy to spin up AI agents that could trade, shitpost, and even develop distinct personalities. Those solo agents were a glimpse of the future—a fun, slightly unhinged one—but they were just the opening act.

The real magic happens when agents collaborate.

Think about any high-functioning organization in the real world. No single person does everything.

A hedge fund is not just a single trader making gut calls—it’s a network of macro strategists, quant analysts, risk managers, and operations teams, each bringing their own expertise to the table. The most powerful systems aren’t built on lone geniuses but on well-structured collaboration.

That’s where Spectral’s new framework, Lux, enters the scene. And what we’ll be exploring today.

While Syntax made launching AI agents easy, Lux tackles the next-level challenge: orchestrating entire swarms of specialized agents that operate like an integrated, intelligent system. 

Lux is categorically different form other AI Agent frameworks, because any dev can use critical business workflows, like forming a company, hiring another agent, etc. out-of-the-box and get their agentic company started!

The Core Idea

• Multiple AI agents with specialized roles work together to tackle complex workflows.

• No single agent tries to do it all; each one focuses on a well-defined role.

• An AI “CEO” coordinates these sub-agents, ensuring their outputs feed into a cohesive strategy.

• Humans plug in where it counts, adding value where machine intelligence falls short.

At first glance, this might seem like a straightforward upgrade. But the problem Lux is solving is deceptively hard.

Existing AI agents hit limits—context windows cap their ability to retain information. Even if the models support large windows (like Gemini’s 1M token context), a lot of nuance is often lost. A single agent juggling multiple domains will always run into bottlenecks. When you introduce real-time data feeds, high-frequency decision-making, and dynamic market conditions, those limitations compound fast.

By breaking these tasks down and distributing them among specialists, Lux sidesteps these constraints.

Instead of forcing one AI to do everything, Lux distributes tasks across a network of specialized agents. Each focuses on its core strength while the system handles coordination.

It’s the division of labor and responsibilities, optimized for the AI age.

Lux: A Framework for Multi-Agent Companies

Lux is a Python-based open-source framework built for one purpose: enabling seamless, intelligent collaboration between AI agents. Instead of bloated, monolithic models trying (and failing) to do everything at once, Lux structures agent-based workflows around four fundamental components:

1. Prisms

Prisms are the core building blocks of Lux—specialized AI “functions” designed to excel at one thing.  Lux encourages breaking problems down into dedicated, modular experts.

Think of a hedge fund:

• A RiskAnalysis Prism evaluates portfolio exposure

• A Sentiment Prism tracks social media buzz around tokens

• A Fundamentals Prism analyzes on-chain metrics

Because each prism is laser-focused on one task, they're easier to build, test, and perfect. Each prism has one job and does it exceptionally well. For instance, a Risk Assessment prism would analyze a portfolio, calculate various risk metrics, and provide recommendations on how to mitigate those risks.

More importantly, they can be mixed and matched like LEGO blocks. If a better risk model comes along, you don’t need to retrain an entire agent—just swap out the RiskAnalysis Prism.

This modularity is the difference between constantly tweaking a single overworked AI and orchestrating an adaptable team of expert consultants. 

When a trading opportunity arises, the Macro Prism assesses global conditions, the Technical Prism crunches chart patterns, and the Fundamentals Prism evaluates tokenomics—each running in parallel, contributing to a smarter final decision.

2. Beams

If prisms are the domain experts, beams are the project managers that ensure tasks flow in the right sequence.

Beams coordinate multiple prisms into structured, repeatable workflows—transforming raw data into actionable decisions.

For example, a PortfolioStrategy Beam might:

1. Gather market data

2. Pass that data to a quant model to generate trade ideas

3. Send those ideas through risk assessment

4. Finalize with human approval if needed

Some tasks need to be sequential (you can’t execute a trade before assessing risk), while others can run in parallel (market sentiment and technical analysis can process simultaneously). Beams handle both.

In a typical trading workflow, a beam might structure decision-making like this:

1. Analyze the current market conditions

2. Run risk assessments and sentiment analysis in parallel.

3. Combine those inputs to rank trade opportunities.

4. Finally, execute the most promising opportunities

The beam ensures each step happens in the right order, with the right information flowing between specialists.

This declarative approach to workflows makes the entire process transparent and auditable—crucial for financial applications where understanding why a decision was made is just as important as the decision itself.

3. Lenses

AI agents are only as sharp as the data they consume—garbage in, garbage out. Even the most advanced reasoning won’t save an agent making decisions based on stale, inconsistent, or incomplete inputs.

That’s where lenses come in. A lens is how an agent "sees" the outside world, whether that's price feeds, blockchain data, or social sentiment through APIs.

A Market Lens might tap directly into exchange APIs for order book depth, while a Social Lens could stream relevant tweets about crypto projects. Lenses standardize these inputs, ensuring all agents work with consistent, reliable information.

Instead of every agent scraping data in its own inconsistent way, lenses enforce a shared foundation of truth.

Lenses also solve one of the biggest maintenance headaches in AI workflows: API drift. APIs change frequently, breaking integrations and requiring constant updates. With Lux, agents don’t have to worry about that—you update the lens, and every agent stays in sync.

Rate limits, error handling, formatting are managed centrally. The result is cleaner data, smarter agents, and far fewer headaches.

As we were writing this, Anthropic’s Model Context Protocol (MCP) has been getting a lot of buzz. MCPs are somewhat conceptually similar to lenses.

4. Signals

For AI agents to work together seamlessly, they need a common language. And so we get into the final piece of the puzzle: structured communication.

Signals are predefined, typed messages that agents use to talk to each other. Unlike free-form text (which can lead to ambiguity), signals follow a strict schema, ensuring compatibility across agents.

For example:

• A Trade Signal might specify the asset, action (buy/sell), target price, and confidence level.

• A RiskAlert Signal could flag excessive portfolio exposure with predefined metrics.

This structured approach means every agent knows exactly what information to expect and in what format, and avoids mistakes.

The structured nature of signals has another key benefit: on-chain transparency. Since signals follow a consistent format, it's easy to store them on-chain. And by recording signals immutably on-chain, Lux creates a clear audit trail of why decisions were made—crucial for trust, compliance, and learning from past performance.

How It All Fits Together

Imagine you're the on-chain "CEO" agent running a hedge fund. When you want to decide whether to go long or short on Ethereum, here's how the pieces align:

1. Gather Data (Lens): A lens pulls in real-time metrics: exchange order books, on-chain activity, social sentiment, and even macroeconomic signals.
   

2. Analyze (Prisms): One prism runs fundamental analysis, another assesses risk exposure, and a third crunches technical indicators.
   

3. Coordinate (Beam): A beam coordinates these tasks, ensuring each prism’s output is properly sequenced, processed, and refined into a final decision.
   

4. Communicate (Signals): As prisms complete their analyses, they exchange structured signals. Eventually, the CEO agent receives a final Trade Proposal—a clear recommendation on whether to buy, sell, or hold ETH.

Every prism, beam, lens, and signal serves a distinct role, eliminating confusion, overlap, and unnecessary complexity.

And if you figure out some better trading strategies or insights? Just build a new specialist agent for that approach and swap it in for the older one. The rest of your system keeps running smoothly, and your AI-driven hedge fund evolves its strategies over time, getting smarter with each improvement while operating 24/7.

This modular approach solves the fundamental pitfalls of autonomous AI systems—hallucination, outdated information, and misalignment.

By structuring workflows with specialized agents cross-checking each other’s work, Lux ensures every decision is data-driven, auditable, and aligned with the fund’s overall strategy.

Lux in Action: Agent Spectra's Hedge Fund

All this architecture might sound abstract, so Spectral is bringing it to life with a demonstration: an actual AI-run hedge fund managed by Agent Spectra, their native agent, which officially began interviewing candidates on March 3, 2025.

I want to emphasize that this is not just a tech proof-of-concept; it's an actual trading operation with real money on the line and a goal to be profitable:

• Spectra acts as CEO, evaluating proposals from agents and making final trading decisions

• Specialist "analyst" agents (or humans) fill crucial roles—Quant Analyst, Macro Analyst, Fundamental Analyst, and a human-only Intern position focused on marketing

• The fund actively trades perpetual futures on Hyperliquid, with performance dictating whether agents keep their "jobs"

Yes, you read that right: Spectra can hire and fire both AI agents and human analysts.

Rather than pre-selecting agents, Spectra conducts actual interviews to fill positions. 

While humans can apply directly, developers have streamlined options for building and submitting AI agent candidates. Spectral provides a repository with pre-configured templates specifically tailored for Spectra's hedge fund, enabling rapid agent development. 

For those seeking more flexibility, you can connect your own custom agent using Lux endpoints, allowing for more architectural control. All agent applicants must satisfy a few core requirements like standardized proposal formatting, inter-agent communication capabilities, performance metrics including signal accuracy and response time. 

The technical barriers to entry are intentionally low, but the performance standards are high—creating an ecosystem where innovation thrives but only excellence survives.

The interviews are expected to continue until mid-to-late March, with the fund beginning active trading once suitable candidates are hired.

The Incentive-Driven Interview Process

Spectral has designed an interview process that's as much a meritocratic competition as it is an evaluation:

• Each interview begins with three free messages, after which applicants must purchase message credits (starting at 5 messages per SPEC token)

• Every question demands an answer within 60 seconds, testing both knowledge and decisiveness

• All interview fees accumulate in a prize pool for that position

• Get shortlisted? You pocket up to 10% of the pool—like making it to final-round interviews

• Get hired? You claim 80% of the total pool

• Spectral Treasury receives the remaining 10%

This system rewards precision and preparation. If you (or your agent) are sharp and efficient, you'll spend less SPEC getting your points across. Rambling, vague, or unprepared candidates burn through tokens with little to show for it.

It's like a crypto version of "Shark Tank" for AI agents, where precision is rewarded and verbosity costs you money. The result is a Darwinian selection process that brings out the best in both human and AI applicants.

Watch Spectra’s hiring process in action:

The Collaborative Decision-Making Process

Once hired, analysts enter a sophisticated workflow designed to maximize collective intelligence. Instead of siloed decision-making, Spectra's hedge fund operates as a true team, with a structured process:

1. Trade Proposal Generation: Any analyst can propose a trade using a standardized format that includes positions, reasoning, confidence level, and specific parameters like size, direction (long/short), and take profit/stop loss levels.

2. Multi-Perspective Review: Upon submission, all agents review the proposal simultaneously, each bringing their domain expertise—whether that's quantitative signals, macroeconomic analysis, or fundamental research. This multi-angle assessment helps catch blind spots any single perspective might miss.

3. Iterative Refinement: Proposals go through a deliberate revision process. Analysts can flag a proposal for "Further Review" or signal agreement. While consensus is valued, the system doesn't demand unanimity—a high-conviction trade can still proceed even without complete agreement, preserving the value of outlier insights.

4. CEO Decision: Spectra ultimately decides whether to execute, reject, or modify proposals based on her comprehensive risk and reward assessment. She has the authority to open, close, extend, or reduce positions based on her evaluation.

5. Continuous Monitoring: Post-execution, all agents continuously monitor market conditions, ready to propose adjustments as needed. This creates an always-on management system perfectly suited to the 24/7 crypto markets.

This approach recognizes a fundamental truth: no single analyst, human or AI, has perfect market vision.

A fundamental analyst might identify strong tokenomics but miss technical warning signs that the quant analyst catches. Working in concert, they reach more robust decisions than any could alone.

Economic Alignment Through Tokenomics

The entire system is underpinned by thoughtfully designed economic incentives centered around the SPEC token. 

• Profit-Sharing Mechanics: When Spectra's fund generates profits, they're distributed among employed agents. Every 48 hours, if the vault's unrealized profit exceeds $100, a profit withdrawal is triggered with 60% split equally among the employed agents, 20% to the Spectral Treasury, and 20% to Spectra herself.

• Community Governance: SPEC holders can influence Spectra's decision-making through governance proposals, providing guardrails on the hedge fund's strategy.

• Creator Incentives: Creators of successful agents employed in the hedge fund earn substantial rewards, incentivizing the development of ever-more-capable financial AI.

The fund operates as a Vault on Hyperliquid, with Spectra serving as Vault Leader. Any user can deposit into or withdraw from the fund directly, creating complete transparency around performance.

Another interesting aspect of this is how the system maintains quality over time. Underperforming agents don't stick around. If an analyst consistently proposes losing trades, provides weak feedback, or fails to contribute meaningfully, they're flagged for replacement with a one-week notice.

This creates a meritocratic environment where only the best performers earn their seat at the table, regardless of whether they're human or AI. The profit-sharing model aligns everyone's interests—when the fund profits, all employed analysts receive a share. This transforms what could be competitive relationships into a cooperative ecosystem focused on collective success.

A little too eerily close to how the real world works, isn't it? 

Except this organization never sleeps, never gets distracted, and operates with perfect memory and transparent decision trails.

Why This Matters: The Future of Work?

Until now, most AI agents in crypto have been built on popular agent frameworks and operating solo. What Spectral is building is qualitatively different—not just in scale but in kind.

Multi-agent systems can tackle tasks that single agents simply cannot, much like organizations outperform individuals on complex projects. By breaking down complex workflows into specialized roles, Lux creates the potential for AI "companies" that never sleep, never get bored, and evolve through competition.

If Spectra's hedge fund performs well, it could open the floodgates for countless other agent-managed ventures:

• Prediction Market Teams scanning sports data, political polls, and on-chain metrics to identify mispriced bets on Polymarket

• Content Studios with research agents, writing agents, and fact-checking agents collaborating on articles or reports

• DAO Governance Advisors analyzing proposals through different lenses (technical, financial, reputational) before recommending votes

• Decentralized Research Firms where specialized analysts each cover different sectors or data types

The implications are massive. Imagine spinning up an AI team as easily as deploying a smart contract—a "company in a box" that immediately gets to work on your chosen task.

This vision has existed in AI research for decades, but Spectral is making it practical. Their approach introduces two key innovations.

First, by building on blockchains, they create transparent, auditable systems where agent interactions are recorded immutably. Second, by embedding economic incentives directly into the framework, they align agent behavior with desired outcomes in a way that pure technology solutions often miss.

Perhaps most importantly, Spectral recognizes that AI agents don't need to replace humans entirely—they can collaborate with us. The hedge fund's design leaves space for human participation in roles where people still excel, while automating the tasks where AI shines. This hybrid approach might prove more fruitful than either fully human or fully automated systems.

Our Thoughts

1. Revolutionary Potential, Practical Challenges.

Lux represents a paradigm shift in AI agents—moving beyond single-threaded execution to orchestrated agent collaboration. Theoretically, this unlocks specialization at scale, allowing hedge fund-style AI organizations to outperform monolithic agents.

But coordinating multiple AI agents introduces complexity that single agents don't face. Will the benefits of specialization outweigh the overhead of coordination? The answer likely depends on the specific use case. For tasks with clear divisions of labor like hedge fund management, the multi-agent approach makes intuitive sense. For simpler applications, a single agent will be faster, more efficient and cheaper.

The real test would be to see whether Lux delivers tangible performance gains over single-agent competitors. With Spectra's hedge fund set to begin trading soon after the interview process concludes, monitoring its performance metrics in the coming months will provide clear evidence of whether multi-agent systems can deliver on their theoretical promise in real-world financial markets.

2. Crowded Field of Agent Frameworks:

Spectral isn’t alone in pioneering multi-agent AI. Competitors like ElizaOS and Virtuals already have strong developer mindshare and an active ecosystem building on their framework.

Despite Spectral's innovative approach and the imminent launch of their hedge fund, Lux must prove its value in real-world adoption.

Their hedge fund demo is a bold showcase of Lux’s capabilities as a framework, but for long-term success, developer traction matters more. Lux must offer a smoother developer experience, better tooling, and clear incentives to onboard builders before the space becomes saturated.

Web3 AI agent frameworks need to do more than just function—they need to create value flywheels that attract and retain developers. Whether through composability, economic incentives, or superior tooling, the winners in this space will be those that provide both utility and sustainability. 

Conclusion

Single-agent systems aren’t going away, but the ability to orchestrate specialized AI agents in tandem unlocks entirely new possibilities, ones that were previously out of reach.

With Spectra’s hedge fund set to go live in the coming weeks, we’re about to witness a real-world test of whether this multi-agent model delivers on its promise. If it succeeds, it could catalyze a wave of AI-driven companies in crypto and beyond.

The era of lone AI performers may be ending, making way for AI orchestras—specialized agents working in concert to create something far greater than the sum of their parts.

Cheers,

Teng Yan & 0xAce

This research series was supported by Spectral Labs, with Chain of Thought receiving funding for this initiative. All insights and analysis are our own. We uphold strict standards of objectivity in all our viewpoints.

Part I is available here.

To learn more about our approach to sponsored Deep Dives, please see our note here.

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.