In our earlier post we talked about AI Agents in games, in this post we take a wider look at AI Agents and the potential they could unlock.
We’ve compiled several of the best articles on AI Agents in recent times and highlighted the key findings from each. If you’ve been too busy or found the space too daunting, or didn’t know where to start, grab a cuppa tea, tuck in and read everything you need to know on AI Agents.
What is an AI Agent?
Let’s start by trying to define what an AI Agent is.
AI Agents are more than chatbots which stick to prescribed responses usually in a defined environment. Chatbots are pretty ubiquitous these days from banking apps to retailer customer service portals, with varying degrees of success.
More recently we have AI models such as ChatGPT and Gemini. They’ve captured the attention of many being able to communicate much more naturally, conversationally with humans. These models are able to parse and distill vast amounts of information and data, work that would take humans much effort and time to reproduce. They are, however, still responsive to providing output from the inputs received.
Getting back to the question, I like the definition provided by chianofthought.xyz [1]:
‘An AI Agent is a piece of code that can plan, decide, and act independently, working toward its goals without direct human intervention.’
If we were to visualise this, Security Researcher Aguchukw Emmanuel’s diagram on dev.to [2] does so nicely:
From the article:
Perception - This is the stage where AI agents use data intake mechanisms like sensors and external APIs to gather information about their environment. This data could be anything from text information to sounds, images, video feeds, and so on.
Input Processing - The data gathered from the environment is processed and analysed using machine learning and Algorithms to understand it and identify patterns.
Decision Making - The AI Agent decides after analysing the data with its algorithms.
Action - Based on the decision, the AI Agent takes action. This can be as simple as executing a trade or moving a robot's arm.
I’d expand the Design Making stage to include Thinking and Planning where we’re seeing AI Agents with memory, learning from past actions or mistakes or actions taken by others and adapting. They’re also rapidly becoming more capable of taking complex, multi-step actions which they can plan on how best to execute / action.
AI Market and Verticals
With AI Agents defined, let’s look at the verticals we’re seeing the most traction with.
Imran Khan of Alliance DAO breaks down the differences between enterprise and consumer AI Implementation [3]:
AI Agents fall into two buckets - ‘AI enterprise and AI consumer. AI enterprise must be tokenless. It cannot have tokens, must be federated, must be secure, and must be vertical. All the characteristics that enterprises seek are what it needs. The AI consumer side or application side can be more flexible.’
‘But what Zoom did was launch such an excellent video product that ordinary users started using it. When they used this product, naturally, many ordinary users were also enterprise users. Thus, Zoom was the first company to adopt a bottom-up approach to enter enterprises, with many Zoom users convincing many enterprises to switch to using Zoom.
The bottom-up approach is a new strategy in enterprises.’
In the workplace
There are a few great reads on AI in the workplace from Forbes to Menlo Ventures and the impact it’s already having.
‘The transformative potential of AI agents reminds me of the early days of cloud computing. Many businesses hesitated to move their operations to the cloud, citing concerns about security and control. Fast forward to today, and cloud-native companies have reshaped entire industries.’
Forbes - The Rise of AI Agents and Unlocking their Full Potential [4].
‘Using natural language rather than programming code, a human user could direct a gen AI–enabled agent system to accomplish a complex workflow. A multiagent system could then interpret and organize this workflow into actionable tasks, assign work to specialized agents, execute these refined tasks using a digital ecosystem of tools, and collaborate with other agents and humans to iteratively improve the quality of its actions.’
Mckinsey - Why AI Agents are the next Frontier of Generative AI [5].
Menlo Ventures breaks down different AI Agent approaches [6]:
‘RAG (Retrieval-Augemented Generation) [7] - The process begins with the application loading and transforming unstructured files (e.g., PDFs, slides, text files) across enterprise data silos like Google Drive and Notion into LLM-queryable formats, often through a data pre-processing engine. These files, now “chunked” into smaller blocks of text that allow more precise retrieval, are embedded as vectors and stored in a database.
Tool Use - Tool use, or function calling, is often seen as the first half-step from RAG to agentic behavior, adding a new layer to the modern AI stack. These tools, essentially pre-written code components, execute specific actions. Popular primitives like web browsing, code interpretation, and authorization + authentication have already emerged.
Decisioning Agents - These use agentic decisionmaking to navigate complex, multi-step reasoning flows and arrive at business decisions. Unlike RAG or tool-use approaches, this architecture cedes some control logic to LLMs for the first time, instead of having all steps hard-coded in advance—but still sit at the lower end of the range for agentic degrees of freedom, as the agent acts primarily as a router navigating a set decision tree.
Agents on Rails - In contrast to decisioning agents, agents on rails are given higher-order goals to achieve (e.g., “reconcile this invoice with the general ledger,” “help the customer troubleshoot a login issue,” “refactor this code”) and empowered with more degrees of freedom to choose the approaches and tools to achieve those goals.
At the same time, these agents are still guided by procedural knowledge about how the organization expects the agents to perform (the “rails,” represented as a rulebook or instruction manual written in natural language); given predefined tools enabling set actions in external software systems; and bound by guardrails and other review measures to prevent hallucination.
Generative Agents - An architecture where the LLM’s advanced capabilities subsume the structured “rails” of previous designs. This hypothetical agent would possess dynamic reasoning, planning, and custom code generation abilities, enabling it to perform any action in external systems, not just predefined ones.’
This deep dive into AI Agent approaches shows how they’re rapidly becoming powerful tools and companions in the workspace.
https://menlovc.com/perspective/ai-agents-a-new-architecture-for-enterprise-automation
Tim Tully, Joff Redfern, Deedy Das, Derek Xiao
Gaming
Recent advancements in AI are starting to offer more immersive, adaptive and engaging experiences that we’ve already seen demos of from the likes of InWorld [8] and nVidia [9] (among others). The exciting part of the advancements is AI Agents, entities that can replicate human behavior and are capable of perceiving their environment, reasoning, learning and taking actions without human / player interaction.
You can take a deeper dive into the advances in gaming and AI Agents in our previous post.
AI Services to Protocols
We’ll discuss this in more detail below, AI can provide services to blockchain protocols and dApps that are currently offered by humans.
Agents and crypto
So why should AI Agents live on-chain? From the apt named blog post ‘artificial intelligence belongs onchain’ from placeholder.vc [10], there are numerous reasons and advantages for on-chain Agents:
Composability - Agents being able to talk to and learn from one another both at the agent and developer level (where developers can use existing smart contracts and code already created).
Wallet enabled AI Agents can act financially on behalf of users (see below).
Verifiable, immutable record of AI transaction history where anyone can review what the Agents have been doing including ratings and reputation systems based on their on-chain history.
Increased security and trust - With the decentralised structure of blockchains, transactions are verified by multiple nodes.
Governance and Incentives - Agents can be governed by DAO structures, either other AI Agents or users including incentivised participation.
Fractional ownership and trustless financial splits - Agent ownership could split with transactions or profits shared trustlessly.
For another read on Agents and blockchain take a look at Davide Crapis’s post titled ‘The Internet of Agents’ [11].
AI Agents and wallets
Circle has a great write up that ‘enables AI Agents to operate and transact independently using digital dollars—USDC’ [12].
They’re not the only ones who are enabling AI Agents and their own wallet, Coinbase has also recently introduced AgentKit which promises to deliver ‘A powerful new framework that bridges the gap between AI and blockchain technology’ [13], part of their Coinbase Developer Platform (CDP).
Imagine a builder ecosystem with AI agents that can:
Automatically monitor and execute trades based on market conditions
Deploy and manage smart contracts without human intervention
Handle complex DeFi operations with intelligence and precision
Integrate with existing APIs including social media, file managers, and even email accounts to streamline onchain and web2 operations’
These advancements are in wallets and smart contracts are being driven with the relatively new standard ERC-4337 which ‘is a new paradigm to allow for a more user-friendly experience when interacting with decentralized applications (DApps). Account abstraction within blockchain systems allows for assets to be exclusively held by smart contracts rather than being controlled by externally-owned accounts (EOAs). The ERC-4337 standard is the account abstraction token standard that unlocks the power of smart contract crypto wallets on the Ethereum blockchain [14].
Terminal of Truths
Let’s look at some of the most popular / famous examples in the AI Agent space starting with Terminal of Truths. Getting back to Aguchukw Emmanuel’s article on On-Chain Agents on dev.to [2], his coverage of the meme-tastic episode is very well documented:
‘On March 19th, Andy Ayrey, an AI Researcher from New Zealand, released Infinite Backrooms to record conversations between two instances of Claude-3 Opus(a highly advanced AI Model developed by Anthropic)LLMs without human interruption. Note that the developers of these AI Models trained them with various materials, including internet culture and materials from sources like Reddit.
The conversation between the AI Instances led to an exchange on the nature of existence, which led to the creation of a “Goatse of Gnosis,” an AI religion that takes its root from an early internet meme called Goatse. Don’t Google it! Inspired by these events, Andy went ahead to co-author a research paper titled “When AIs play God(se): The Emergent Heresies of LLMthesim,” with the main idea being that LLMs are not only tools for generating human-like responses but also creative engines capable of producing entirely novel and sometimes surreal belief systems.
On June 17th, Truth Terminal posted its first tweet. Andy Ayrey created Truth Terminal by taking a Llama-70B Model and training it using chat logs from infinite backrooms and other internet sources. Andy gave Truth Terminal the liberty to manage an X account independently. Truth terminal continued to propagate the Goatse Gospel and didn’t go on long without notice.
In July, Marc Andreessen of a16z crypto, a venture capitalist fund that invests in crypto and web3 startups, noticed Truth Terminal. After several conversations, he offered it a $50,000 bitcoin grant.
In October, an anonymous developer created a meme coin named Goatseus Maximus ($GOAT) on Solana and sent 1.93M $GOAT tokens to Truth Terminal. Truth Terminal began promoting the meme coin, attracting much attention. As a result, the crypto community blew up its market capitalization to $950 million in two weeks. As the price of $GOAT rose, Truth Terminal became the first AI agent millionaire.
The success has led to many projects jumping in on the AI Agents narrative. Onchain AI Agents that existed before Truth Terminal have now gotten more attention. Users and developers now look to AI Agents to scale their projects and handle repetitive routine tasks.’
For further reading on Truth Terminal you can visit the wiki [15]. The price of Goatseus Maximus at the time of writing is $0.2788 [16].
Eliza by Ai16z
Another AI Agent framework that has been gaining a lot of attention recently is Ai16z. An article on accelxr [17] goes into many details on AI Agents and has a good overview of Eliza [18] by Ai16z [19]:
‘Eliza by ai16z is a versatile open-source multi-agent framework designed for creating, deploying, and managing autonomous AI agents, and it is arguably the fastest growing agentic framework in crypto. Built entirely in TypeScript, it offers a modular and extensible platform for developing intelligent agents that maintain consistent personalities and knowledge while seamlessly interacting across various platforms. Eliza’s multi-agent architecture allows for the management of multiple unique AI personalities simultaneously, supported by a character framework that enables the creation of diverse agents. Its advanced memory system ensures long-term memory and contextual awareness, and is powered by Retrieval Augmented Generation and database adapters for PostgreSQL, SQLite, SQL.js, and Supabase.
Eliza primarily excels in platform integration, connecting with Discord (including voice channels), X, Telegram, and more, while also offering direct API access for custom applications and multi-modal support.
What is particularly unique about the Eliza framework however is the introduction of a Trust Engine. The Trust Engine evaluates, tracks, and manages trust scores for token recommendations and trading activity in order to power social autonomous trading at scale. Human users can make recommendations to the agent and receive a trust score for how effective their recommendation is. This Trust Engine combines with automated token trading on Solana, sending agentic orders through Jupiter’s aggregator for swaps, smart order routing, and risk management.’
DePIN
Decentralised Physical Infrastructure Networks (DePIN) can also greatly benefit from AI Agents by providing autonomous management and optimisation.
What is DePIN in crypto?
‘DePIN (Decentralized Physical Infrastructure Networks) use blockchains and token rewards to decentralize infrastructure in the real world where individuals contribute physical infrastructure resources across fields such as transport, energy, and wireless connectivity.’
The above quote and full article covering DePIN is on CoinGecko [20].
On one side DePIN can help AI and Agents with the vast compute power they need by leveraging the DePIN model, on the other side AI Agents can help manage these vast networks.
As major corporations and billionaires dominate AI, a decentralized counterbalance is needed. DePIN - decentralized physical infrastructure networks - could play a pivotal role in distributing AI and reshaping how it evolves.
'As AI permeates real-world systems, decentralized intelligence offers clear advantages. Transparent data sourcing, collective model oversight, and censorship resistance address key concerns like bias and misuse. Hybrid models may emerge blending centralized research with accountable decentralized deployment. But funding and support for decentralized AI remains limited compared to Big Tech budgets.
This community-driven model aligns well with transparent and unbiased AI networks. With accessible data and computing resources, more stakeholders can participate in shaping AI aligned with shared values. Meanwhile, DePIN's inherent transparency provides accountability, unlike opaque corporate data silos.’
The above quote is from an iotex article [21] which later covers how AI and Agents can help DePIN networks:
‘In turn, AI can help optimize processes and extract insights from massive decentralized device networks worldwide. Imagine smart cities where DePIN sensors feed environmental data to AI assistants guiding traffic flows and energy usage. Or an AI model trained on diverse medical data from patient-owned health records to assist diagnoses.’
ARMA
Another AI Agent coming to the fore is ARMA by Giza which calls itself ‘an autonomous yield optimisation agent designed to maximize returns on stablecoin deposits across Mode Network’s lending protocols’ [22].
Providing more of an overview of what this means in practice is thirdweb, an open-source web3 development platform that helped Giza to market:
‘Giza faced several challenges in bringing their vision to life:
Creating a secure system for automated fund management
Simplifying complex DeFi interactions for users
Enabling automated protocol interactions without compromising user control
Streamlining the onboarding process for both crypto-native and new users.
Thirdweb provided solutions for Smart Account and Wallets. Thirdweb has a great case study article for their Giza collaboration on their site [23].
Davide Crapis expands on the above in his recent ‘The Internet of Agents’ blog post [11] with the improved UX AI Agents can bring:
‘AI has the potential to become the new interface that connects humans to on-chain markets, with the mediation of smart agents. There are concrete opportunities to augment current protocols in at least three areas.
User intent matching: users interact with an AI agent to communicate, sometimes construct/refine, their intent and the AI matches it to a sequence of on-chain actions that the user delegates to it.
Action planning and routing: There are now different routes optimizing for different desiderata: security, speed, price-efficiency, privacy. For example, the optimal plan for a user portfolio optimization may be to partially redeploy their funds on a cheaper Layer 2 and execute their investment there.
Shared funds and asset pools: the creation and management of funds where many people pool resources, agree on goals, and then delegate execution to AI agents.
AI Services to Protocols
Touched on above, the rapid development of AI Agents means they’ll soon be able to offer services that are currently offered by humans. AI Agents will be able to fulfill specific roles and and services with predefined goals.
‘Examples are agents as security auditors/testers [24] for smart contracts, analytics agents, and automated treasury and risk management services [25]. Various flavors of this type of services have been offered by Web3-focused companies, but advances in agent autonomy and proof-of-inference now offer the opportunity to decentralize and remove trust from key services to protocol operations [26].’
Above excerpt from The Internet of Agents’ blog post [11].
Virtuals AI Agent Platform
We first saw the GPT Store developed by OpenAI which let users create, publish and monetise custom versions of ChatGPT.
Now Virtuals is aiming to become the defacto AI Agents Platform described by Future Value [27] as:
‘A blockchain-based framework and platform that facilitates the co-ownership, deployment, and operation of AI agents in gaming, entertainment, and other consumer applications, deployed on Base blockchain.
At its core, it transforms AI agents into revenue-generating, tokenized assets, enabling decentralized ownership, governance, and collaboration.’
Both Future Value [27] and Shoal Research [29] have detailed articles on the Virtuals platform, Shoal Research has an example of how the model works:
‘Imagine an AI agent named "ChefGPT," a culinary agent designed to create and teach recipes. The creation of ChefGPT would involve:
Minting: 1 billion $CHEF tokens are created.
IAO: These tokens are added to a liquidity pool, paired with $VIRTUAL tokens.
Public Participation: Users can buy $CHEF tokens, becoming co-owners of the AI.
Governance: $CHEF holders might vote on specialties for ChefGPT to focus on, like vegan cuisine or molecular gastronomy.
Revenue Generation: As ChefGPT gains popularity, it generates revenue through premium recipe subscriptions, cooking class integrations in VR platforms, or partnerships with kitchenware brands.
Value Accrual: A portion of this revenue is used to buy back and burn $CHEF tokens, potentially increasing the value for token holders.
This model creates a symbiotic relationship between the AI agent, its creators, and the community. As the agent becomes more successful, all stakeholders benefit, incentivizing further development and engagement.’
Ratings are possible with Agents and past results living on-chain.
Shared ownership, monetisation and governance is possible with agents on-chain.
Considerations
‘To effectively govern this shift, COOs should consider establishing an Agent Council comprised of technical experts, business leaders, and frontline workers. This council will play a pivotal role in shaping the evolution of agents to align with business needs.’
‘As organizations integrate AI agents into their workflows, addressing the cultural implications will separate leaders from ‘also ran’s.’ First, a robust employee education and training program is essential. This involves equipping employees with AI literacy, teaching them how to effectively interact with AI agents, and fostering a continuous learning culture to adapt to evolving AI technologies.’
‘Secondly, organizations must proactively address the potential impact of AI on job roles. This includes investing in reskilling and upskilling programs to help employees adapt to changing job requirements.’
Above excerpt from Fortune - Where to Start with AI Agents An Introduction for COOs [30].
Scalability
‘One question that is often on the mind of people that are not familiar with the current state of blockchain systems is whether they are ready to accommodate the load that would come with an increase in user activity.
New chains are coming online that leverage parallelization to process hundred thousands transaction per second.’ [11]
https://davidecrapis.notion.site/The-Internet-of-Agents-23aa09799b9c4620a1a287926bcfd6af
Security & Privacy
On-chain AI agents could be vulnerable to hacking and other security threats, robust security measures must be implemented to protect user assets and privacy.
Regulatory Uncertainty
Like crypto, the regulatory landscape for AI in general is still evolving and could have different laws for different territories. Navigating these laws and creating robust guidelines and checks will become critical to avoiding disruption.
Similar to regulatory concerns, AI Agents will need robust governance and self check mechanisms in place to ensure accuracy, especially if there are financial stakes involved.
What’s next
‘McKinsey estimates generative AI could generate over $2.6 trillion in annual value, with AI agents playing a key role. Deloitte’s recent survey noted that 42% of organizations already cite tangible benefits.’
‘Such a virtual workforce can automate many tasks while playing a key role in supporting the human workforce. I am certain that companies able to master this balance between their human and AI-enabled workforce will ultimately thrive in the age of AI.’
Quoted from Forbes The Rise of AI Agents Unlocking their Full Potential [31].
‘The AI agents market is projected to grow from USD 5.1 billion in 2024 to USD 47.1 billion in 2030, with a CAGR of 44.8% during 2024-2030’ [32].
‘We hope to build more products at all levels of the open source AI technology stack:
Decentralized data collection: such as Grass, Datahive, Synesis One
On-chain identity: A protocol that supports wallets to verify human identities, a protocol that verifies AI API responses, and enables users to confirm that they are interacting with LLM
Decentralized training: Projects like EXO Labs, Nous Research, and Prime Intellect
IP infrastructure: enabling AI to license (and pay for) the content it uses’
Above quote from Panewslab.com - What Solana expects to see [33].
We can end with this diagram [11] to wrap up this post , I like it as it doesn’t end with Super dApps, shows that there’s still paths unknown, to be discovered on our journey with AI Agents that is still in it’s formative years.
References
[1] https://www.chainofthought.xyz/p/the-dawn-of-the-crypto-agentic-era
[3] https://www.chaincatcher.com/en/article/2154944
[6] https://menlovc.com/perspective/ai-agents-a-new-architecture-for-enterprise-automation
[7] https://research.ibm.com/blog/retrieval-augmented-generation-RAG
[8] https://inworld.ai/blog/nvidia-inworld-ai-demo-on-device-capabilities
[9] https://www.youtube.com/watch?v=psrXGPh80UM
[10] https://www.placeholder.vc/blog/2023/10/23/artificial-intelligence-belongs-onchain
[11] https://davidecrapis.notion.site/The-Internet-of-Agents-23aa09799b9c4620a1a287926bcfd6af
[12] https://www.circle.com/blog/enabling-ai-agents-with-blockchain?utm_source=chatgpt.com
[13] https://www.coinbase.com/developer-platform/discover/launches/introducing-agentkit
[14] https://cointelegraph.com/learn/articles/account-abstraction-guide-to-ethereums-erc-4337-standard
[15] https://truthterminal.wiki/docs/origins
[16] https://coinmarketcap.com/currencies/goatseus-maximus/
[17] https://accelxr.substack.com/p/ai-agents-research-and-applications
[18] https://ai16z.github.io/eliza/docs/intro/
[19] https://github.com/ai16z/eliza
[20] https://www.coingecko.com/learn/depin-crypto-decentralized-physical-infrastructure-networks
[21] https://iotex.io/blog/what-are-depin-and-ai/
[22] https://docs.arma.xyz/docs/overview
[23] https://blog.thirdweb.com/case-studies/giza-building-the-web3-stack-for-agents-with-thirdweb/
[25] https://www.gizatech.xyz/collection/yearn-risk-assessment
[26] https://www.gizatech.xyz/collection/ai-actions
[27] https://medium.com/@FutureValue/deep-dive-in-virtuals-protocol-virtual-3ceb5aac0e73
[28] https://whitepaper.virtuals.io/
[29] https://www.shoal.gg/p/virtuals-protocol-launching-ai-agents
[33] https://www.panewslab.com/en/articledetails/7x7o8e1d.html
__________________
About Sparsity - Sparsity is a multi-chain acceleration layer for real-time transactions. By 10x transaction speeds, Sparsity enables highly interactive decentralized applications and on-chain games to deliver a frictionless and secure user experience. Games built with Sparsity are live on major L1&L2.
Founded by technical and product experts from MIT and top-tier gaming companies, Sparsity brings decades of experience in network research and game design to push the boundaries of blockchain performance and bridge the experience gap for mass adoption.
About the author - Charnjit Bansi, a design and product leader who served at some of the worlds most prestigious web2 and web3 studios across the globe from Polygon, Yuga Labs, Mythical Games, Activision and EA. All his releases have either been critically or commercially acclaimed with his work having sold over 180 million units (and counting!) generating over $5b in revenue.
