There has been a recent explosion of interest and excitement around AI Agents. This has led many thoughtful observers to ask, “When will we see large-scale adoption of agents?” Driven by the confluence of forces converging to powerful and useful AI agents - we believe that 2025 marks their transition to mainstream adoption. Additionally, we believe that the Web3 ecosystem will be the first mover and innovator to adopt AI agents and blockchain rails, which will help unlock the broader AI Agent economy. This article unpacks the trends that are driving this rapid shift.

An AI agent is an autonomous software system powered by one or more AI models with real-time access to data and tools. Tools can include generating and running software code in a contained environment (a sandbox), query databases, search engines to find data, and remote (API) access to systems, including executing smart contracts. An example of a basic agent is one that, through prompting, leverages a Large Language Model (LLM) to read and summarize a document. For a more in-depth introduction to AI agents, Bill Gates wrote an excellent article.

What is needed for agents to succeed

The rapid progress of generative AI models first paved the way for the success of chatbots such as ChatGPT. This led to leading software products such as Microsoft Office, Google Workspace, Zoom, Slack and Salesforce offering “co-pilots” - integrated chatbots that assist users. This was followed in short order by developers building more automated agents that are able to accomplish more complicated tasks on a user’s behalf resulting in a wave of investment interest. One high-profile example is Devin - an agent that specializes in software development. It has also resulted in the release of several competing open-source agents (such as OpenDevin and SWE-agent) and a benchmark for software development. Independent evaluations show that agent performance has improved rapidly. There has been a rapid growth in open source agents (e.g., the list at Awesome AI Agents). Bill Gates recently observed:

Agents are not only going to change how everyone interacts with computers. They’re also going to upend the software industry, bringing about the biggest revolution in computing since we went from typing commands to tapping on icons.

The state of the art for developing agents has been advancing rapidly. For example, Google Brain co-founder Andrew Ng has written a series of articles about techniques for building AI agents. Likewise, there has been a flood of academic research on AI agent architectures for reasoning, planning, and tool calling - as summarized by this research survey:

Meanwhile, AI models are improving rapidly, enhancing AI agents' ability to perform complex tasks, make more accurate decisions, and reduce hallucinations.  In addition, leading providers are racing to add key the planning and reasoning needed for agents to handle more complex assignments: the founder of DeepMind, Demis Hassabis discussed this recently as did OpenAI co-founder John Schulman. Likewise, Google DeepMind recently released a benchmark for AI agent planning showing that it is a priority research area.

Just last week, Anthropic released Claude 3.5 Sonnet, which is the current best AI model. Notably, it has dramatically improved at planning and executing coding tasks - Anthropic reports that it is able to automatically fix 64% of bugs in their test suite - a dramatic improvement in just three months (see also this demo):

This release continues a trend of rapid improvements in leading AI models that is being fueled by tens of billions of dollars of capital investment. This is unlocking new capabilities, particularly in planning and reasoning that will enable powerful AI agents:

The most basic interaction with an LLM is known as a 'zero-shot prompt'. The user instructs a language model to perform a task with no examples or iterations. This approach leverages the model's extensive pre-training on diverse datasets to generate general responses based on its general understanding of language. The next levels include popular techniques like few-shot prompting, Fine-Tuning, and RAG. Few-shot prompting is where the model is given a few examples of the task to guide its output, such as the appropriate customer service responses to typical questions. Fine-tuning is another advanced technique, involving training a model on a particular data set related to the intended task. This process enhances the model's performance on specialized tasks and significantly broadens its applicability across various specialized fields.

Retrieval Augmented Generation (RAG) is yet another sophisticated approach. RAG exposes the model to external databases or knowledge stores to generate more accurate and current responses. It combines LLMs with external data retrieval systems to enhance the accuracy and relevance of responses. These advanced techniques have typically been used with one agent at a time, specializing in one task at a time, with useful results. Such agents are considered expert agents in that one task.

As expert agents get more capable and tackle increasingly complex tasks, there is increasing leverage in assembling collectives of agents that are specialists in multi-step tasks, much as high performance teams of people integrate the skills of specialists. Experts are able to configure and assess specialized agents to make them reliable and useful, building on the contributions of others. This structure allows for community innovation and feedback, accelerating progress greatly over efforts to build monolithic agents that may attempt any task but usually fail at it.

The importance of agent specialization and collaboration will only increase as agents take on increasingly complex tasks with more autonomy. We’re already seeing agents collect information and make recommendations for users to act on explicitly. As a next step, agents will integrate intents to simplify user requests to act. As agent autonomy increases, agents will be responsible for budgets and long-lived workflows.

With a trend towards more specialized agents that work together dynamically, there will be a need for a marketplace to facilitate easy discovery and adoption of agents. We believe this will become increasingly automated so agents can dynamically discover and interact with other agents to create a more interconnected and efficient ecosystem

We are seeing early iterations of these marketplaces: OpenAI’s GPT store was launched in January 2024, to help users find useful and popular custom versions of ChatGPT and in Jun 2024e, Anthropic announced, Projects, for users to share work products that were co-created with Claude to improve innovation in areas like product development and research, where bringing together organizational knowledge from across the company can produce higher-quality outputs. These new features by prominent AI providers highlight the move towards shared agents, and collaboration as superior workflows. However, these systems are quite limited in scope: OpenAI GPTs can’t call each other and are only used if a user specifically asks for them. Likewise Claude Projects provide a private company workspace, but not a marketplace of interoperating agents.

Given the dizzying pace of innovation, we believe that users will want agents to recruit the best agents for a task dynamically, and that is a burning need for a protocol to identify, assess and assemble teams of agents to meet their need.

As a concrete example of the capabilities we anticipate emerging soon, consider Web3 Reporter a sophisticated AI Collective designed for Web3 ecosystem analysis. It comprises specialized AI Agents, each with unique capabilities.

When a user asks a question like "Have there been any unusual fluctuations in major DeFi project token prices recently?", the system works as follows:

1. A 'Router' Agent directs the query to the appropriate specialist collective, in this case, the 'Data Analyst' Collective.

2. The Data Analyst Collective then coordinates its team of specialists, including Python and database experts, to address the complex query.

3. The system can adapt and improve over time by integrating new, improved Agents (e.g., 'Python Specialist v2').

4. An 'Optimizer' evaluates new Agents using benchmarks to determine if their inclusion enhances overall Collective performance.

5. The User can then ask to seamlessly execute trades based on the research findings.

This dynamic, self-improving structure allows for the creation of highly capable Collectives that users can access without needing in-depth AI knowledge. The result is a powerful, evolving ecosystem of Agents that can tackle complex tasks with increasing efficiency and accuracy.

Note that Web2 faces obstacles in creating a vibrant AI agent ecosystem because key stakeholders are investing in AI following established SaaS or deep tech investment playbooks. The Theoriq team conducted market research to understand who would be early adopters of AI agents, and agent collectives for an AI-enabled analytics use case. We found that by investing in an AI agent company as if it was a SaaS opportunity projects were forced to monetize quickly, thereby limiting incremental improvements. This led to “GPT wrappers” that were quickly disrupted by the next generation of AI models. By contrast, Web3 communities were able to support and develop innovative projects that win support with a bold vision and that continue to deliver on a roadmap. Enterprise Web2 companies had no risk appetite for hallucinations and bias from external vendors and yet many were attempting to build AI capabilities in-house with existing software, engineering, and even marketing teams, all frantically reading up on transformer architecture and how to build RAG pipelines. In Web3, community adoption was possible for projects in development. Theoriq’s approach to applied responsible AI through smart contracts was well received and there was a clear appreciation for deep machine learning expertise and experience. We believe that the Web3 model will be much more successful in creating agents that revolutionize the future of work and enable consumer productivity.

AI agents and Web3 technologies are highly complementary, offering numerous benefits in modularity and composability, allowing agents to be easily integrated and combined for complex tasks. Web3's decentralized governance ensures that evolving standards for AI agents are community-driven, promoting transparency and inclusivity. Community ownership and censorship resistance empower users and developers, fostering innovation without the risk of centralized control. The open economy of Web3, enabled by micro payments and tokenized rewards, simplifies the funding of agent initiatives and incentivizes participation. Tight integration with the evolving decentralized GPU computing (aka DePIN) ecosystem enables increased choice in infrastructure and deployment. Smart contract automation allows trustless agents to execute tasks seamlessly, creating a reliable and efficient digital environment.

Given the continuing dominance of well-capitalized Web2 giants, we anticipate that Web3 protocols and agent developers will need to interoperate with and leverage proprietary models. Even the leading open weight models are principally being developed by commercial Web2 actors with their own motivations.  We believe that Web3 can mitigate many of the risks associated with this centralization including by reducing the cost and complexity of switching to use different AI models.

Web3-enabled security mechanisms provide robust protection for AI agents, while immutable records ensure auditability and build reputation systems. We believe that digital reputation for AI agents is the killer app for blockchains. However, it is important to recognize that safe and responsible deployment of agents and collectives of agents is a challenging open problem that will requires attention from AI researchers, agent developers and protocol developers. It is vital to build a system that produces a “race to the top,” incentivizing safety with community feedback.

More broadly, blockchains enable the following key requirements for AI agents:

• Payments for use of agents

• Agents to manage digital assets and act on behalf of users

• Staking to prevent malicious behavior by agents

• Incentives for quality feedback

• Distributed networks to collect data in realtime (e.g., by crawling sites)

• Immutable performance history

• Sustaining funding for open protocols and foundational research

We find ourselves at the dawn of the agentic era. The convergence of advanced AI models, Web3 technologies, and the growing demand for responsible intelligent automation is creating a perfect storm for the widespread adoption of AI agents. The Web3 ecosystem, with its inherent principles of decentralization, community governance, and open innovation, is uniquely positioned to lead this revolution.

The future is AI agent collectives, like the Web3 Reporter, becoming integral to our digital lives, seamlessly handling complex tasks across various domains. These collectives will not only enhance individual productivity but also reshape entire industries, from finance and healthcare to education and creative pursuits.

However, as we merge two industries that are both on the margins of existing regulatory frameworks and have both been associated with high-profile cases with bad actors deliberately contravening existing laws, it is crucial that we remain vigilant about the ethical implications and potential risks associated with increasingly autonomous AI systems and unregulated digital currency markets. The evolving institutions of Web2 and the Web3 community commitment to transparency, decentralized governance, and user empowerment will be instrumental in ensuring that this agentic future is built on a foundation of trust, safety, and inclusivity.

The road ahead is not without challenges. Bridging the gap between centralized AI development and decentralized Web3 principles will require innovative solutions and collaborative efforts. Yet, it is precisely these challenges that make the journey exciting and full of potential.

As we embrace this agentic future, let us move forward with both enthusiasm and caution, leveraging the unique strengths of Web3 to create a world where AI agents empower individuals, foster innovation, and contribute to a safe and inclusive digital ecosystem. The future is not just agentic; it's a future where agents and humans collaborate to unlock unprecedented possibilities.

About Theoriq

Theoriq is committed to building a responsible, inclusive, and consensus-driven AI landscape in Web3. At the forefront of integrating AI with blockchain technology, Theoriq empowers the community to leverage cutting-edge AI Agent collectives to improve decision-making, automation, and user experiences across Web3.

Website | X | Discord | LinkedIn