Autonomous Agents in Autonomous Worlds
A virtuous cycle.
Autonomous Agents?
In other articles we have talked at length about the intersection of AI and blockchain, and more specifically about the role of autonomous agents in our future world and DeFi.
These are software programmes that think for themselves via LLMs and can execute tasks.
The focus of this article is the impact of such agents in Autonomous Worlds. Not only will agents prosper in these synthetic training environments, but they are actually necessary for Autonomous Worlds to reach their full potential.
Autonomous Worlds?
Autonomous Worlds are one of the most exciting developments in the metaverse. In their basic form, they are persistent open-source environments which have hard-coded digital physics. They invite independent developers to build within and to create experiences upon.
I have written extensively about emerging trends in fully on chain games. Although primitive right now, some of these games are being iterated on by developers to slowly transform them in to Autonomous Worlds. Notable examples include Sky Strife, Mithraeum, Citadel, Conquest, Dope Wars, Topology, Downstream, The Cursed Machine and Primordium.
Having grown up on historical strategy games, I am most excited for the Realms Autonomous World (RAW, aka Realms.World). Not only does it have the grandest vision but it has also attracted the largest set of independent developers through its Frontinus House grants system. As such, it embodies the essence of an open Autonomous World.
RAW is primarily focused on two distinct games that are in development: Eternum, a grand strategy (4X) game where players build kingdoms, and Adventurers, a role playing game (RPG) where players embark on quests.
Consequently, two very separate gaming communities, 4Xers and RPGers, will collide together within the same environment for the first time in gaming history. Extremely novel and unpredictable gameplay loops will emerge from their interactions.
Many more games are in development within the RAW which has just announced a Layer3 that will allow for very cheap and fast transactions.
Autonomous Agent State of the Art
2023 was quite a year.
In April, Generative Agents: Interactive Simulacra of Human Behavior was published, describing the virtual village of Smallville, a game-engine powered simulation of 25 autonomous AI agents. Each was bestowed with a paragraph outlining their identity and then set free to form intricate relationships with others. These agents could interact through ChatGPTs like dialogues, creating a dynamic social environment.
Agents showcased self-awareness, memory retrieval, plan generation, reactions to stimuli, and reflective thinking. Their AI integration enabled them to exhibit behaviours mirroring real humans, such as organising their days, forming and nurturing relationships, and even planning social events like parties.
An intriguing aspect of this simulation was the agents’ ability to engage in self-reflection about their identities and goals, leading to creative decisions based on synthesised experiences. For instance, Maria, chose a music-themed gift for Wolfgang by contemplating his interests.
The social dynamics within Smallville were emergent and complex, characterised by the diffusion of information about various events and the spontaneous formation of new relationships among the agents. This simulation represented a groundbreaking blend of AI and social simulation, mirrored by other frameworks, offering insights into the potential for autonomous agents to create rich, human-like social environments.
In May, Voyager was published, a pioneering LLM-powered agent unleashed within the world of Minecraft. This groundbreaking study redefined gaming autonomy: a self-directed learning curriculum maximising exploration, an ever-expanding repertoire of complex, executable skills, and an iterative prompting mechanism that evolves through environmental feedback and self-reflection. A quantum leap in the potential of AI-driven gaming experiences.
In June, NVIDIA released this breathtaking demo of the integration of real-time AI within a video game, specifically focusing on the interaction between human players and non-player characters (NPCs), all being done locally. Traditionally, NPCs have limited pre-programmed dialogue, but with the advent of generative AI these characters can now respond with real-time dialogue. This advancement significantly enhances gaming experiences, making them more immersive.
In July, the Olas network fair launched via a liquidity bootstrapping pool. This is a chain-agnostic generalised framework for off-chain autonomous agents designed to interact on-chain. It allows for the creation of co-owned, co-governed multi agent systems powered by blockchain.
There’s a general “evolutionary” pressure of AI more broadly (specifically LLMs) to become more agentic and so working out ways in which we can co-own them is essential.
The Cambrian explosion of experimentation in this space is devoid of any blockchain based options, which is what makes OLAS so exciting.
In August, Communicative Agents for Software Development was published, describing an approach with individual GPTs split into specialised “brains”. This concept was applied in an imaginative experiment where AI agents were tasked to form a virtual game development company.
Each agent was assigned specific roles, such as CEO, CTO, designers, and coders, mirroring the structure of a real-life organisation. These roles included managing teams responsible for coding, testing, and communication, with a chat chain established for seamless team interaction.
In this setup, programmers followed instructions from the CTO, while testers focused on identifying bugs and proposing solutions to reduce errors. This collaborative process led to the rapid creation of a 5-in-a-row game. Initially developed without an interface, designers later iterated on the design aspect.
Impressively, the entire development process, from conception to interface design, was completed in less than seven minutes for under one dollar, showcasing the efficiency and potential of AI collaboration in game development.
For a deeper description of how “self-owning” agents may make decisions, build relationships and evolve in unpredictable ways that lead to emergent behaviours, I recommend this excellent article by Tim Cotten.
Autonomous Agents <> Autonomous World Reciprocity
True autonomy is achieved via property rights, markets and privacy:
Charlie has recently published a four part series that delves into property rights and the implications they have for the metaverse. We have written at length about open vs closed artificial intelligence, and his larger thesis supported many of the concepts found there.
If you can’t pay for your resources → you can’t have autonomy. In summary, open metaverses offer an environment where control over digital assets and identities, community-driven governance, and economic models supportive of user sovereignty converge to provide true and complete autonomy for individuals. This autonomy is crucial for protecting against the potential negative impacts of surveillance and control in digital spaces.
Blockchains enshrine property rights and facilitate storytelling:
Blockchain technology plays a critical role in the above. Their immutable ledgers record ownership over value and information whilst offering robust security features to safeguard user data and privacy. Tokens shape participation and innovation through incentive structures underpinned by traceable transactions. NFTs create digital scarcity and verifiable ownership of virtual assets. They can act as collaborative CC0 storytelling primitives and co-governed narrative devices.
Integration of blockchain-based digital identities and reputation primitives will enhance user control over personal data and privacy. Consequently, blockchains foster a fair, transparent and empowering metaverse. They ensure trust.
Autonomous Worlds cannot exist without blockchains and a common state:
An Autonomous World’s success hinges on the ability of users to have a say in its governance and to exercise a degree of autonomy. Through these mechanisms, they foster innovation by providing a transparent environment for developers and creators. Ownership and monetisation of what has been created. They empower users through their control of their data and their digital experiences. Equitable compensation to those that deliver fulfilment.
Autonomous worlds cannot work without: (a) smart contracts to automate and enforce agreements, thus supporting and securing digital property rights; (b) decentralised applications for secure and self-sovereign digital experiences; (c) DAOs that enable community-driven governance through blockchain-powered voting and reputation systems; (d) the composability of blockchain assets that allows for the blending on-chain environments into larger worlds; and (e) a common state in order to move forward and manage risk — see social media today and decay of “truth”. Problems around optimising computation, abstracting complexity, improving scalability, and cross-chain interactions will be solved.
Autonomous Worlds represent the next generation of gaming:
The use of blockchain-powered digital physics will create forces that are aligned with or against each other. Through natural on-chain events, individuals and entities will automatically try and gain an advantage over one another. There is no need to seed the world with evil bosses or fulfilling events; they will evolve naturally. If you start hording a scarce resource, someone is going to come gunning for you. These are the types of games that you can play forever.
Persistent worlds make you care more about what you do; all actions affect all future playthroughs. They foster an environment within which players create things they truly cherish. That emotive feeling when something that costs meaningful time and tangible value is taken away from you is unparalleled. World-renowned video game designer Hideo Kojima has spoken about this concept since 1998, something referred to as “project raw” (quite the coincidence!). This is a game which you play for years and when you die that’s it.
Autonomous Worlds are supercharged by AI Agents:
This is about so much more than just using AI for generative world-building. As the “Communicative Agents’’ paper illustrated, realistic paths exist to creating unprecedentedly deep gaming experiences. This has been functionally impossible until now due to the human constraints of massive budgets, development times, and expanding scope. These were a requisite for handcrafting all aspects of modern games that aim to continually surpass their beloved predecessors. This is exemplified by GTA 6… it’s been ten years since GTA5!
The tools are already available. Emergent gaming interactions are now possible with agentic LLM systems — they result in seamless and intuitive human-AI interaction. Agents can already participate in negotiation, tactical coordination and strategic reasoning at a human level. Models have been created that facilitate collective, decentralised narrative-building linked to a blockchain; woven from the same fabric as AI agents. Essentially, we can already rely on agents to create everlasting worlds, something that humans could never maintain.
To realise the true potential, agents will operate in worlds where they have autonomy:
Within Autonomous Worlds, the provenance of complex social dynamics is guaranteed, so agents can trust that their future interactions have a probabilistic and predictable nature. It allows them to put their hard-earned value at risk for calculated rewards. They will thrive in native digital asset economies, whether they be resource-based or reputation-based assets earned through competition.
The interplay of centralisation and decentralisation in power dynamics impacts autonomy and collective identity, as Charlie has detailed. Composable treaties for player-agent agreements, hidden but provable information for secret roles, trustless enforcement of gameplay rules, mechanics and asset ownership, fully inspectable game states — the underpinnings of property rights and privacy — these are things agents will never rely on from centralised authorities. They are attracted to their decentralised counterparts, where they are granted true autonomy.
The assured migration of these agents to these realms will expand design spaces and lower barriers for human players to access and engage with autonomous on-chain game worlds. Autonomous Agents are essential for Autonomous Worlds.
I started by describing Autonomous Worlds as metaverses. There is a near future where these metaverses are populated not just by you and I, but by a wonderful eclectic mix of AI born autonomous citizens. To fully appreciate their impact, I invite you to read this article by Agent 23.
We believe that one of the first places such interactions will be observed will be within games such as Eternum.
My gratitude extends to Aaron Wright, ChrisF, Loaf, Secretive, Nick Butts, David Minarsch, pet3rpan, Rob Morris, Tim Cotten and everyone on the ID Theory team for reviewing this article and providing critical feedback.
ID Theory may hold positions in some of the projects discussed in the above. This article is strictly for informational and educational purposes only. It does not in any way constitute an offer or solicitation of an offer to buy or sell any investment or cryptoassets discussed herein. Always perform your own research and conduct independent due diligence prior to making any investment decisions.
Interested in partnering with ID Theory or building something special? Get in touch through our website or at info@idtheory.io.
