Agents Gone Wild

Welcome to Agents Gone Wild! Here, we’ll discuss multi-user artificial intelligence. That is, AI agents interacting for multiple users across personal and organizational boundaries.

I’m Joe, an AI-focused software developer and tech/insurance lawyer. I’m the founder of Grounds, which builds AI systems for litigation, negotiations, and dispute resolution, and I’m a partner at Lopez, Bark & Schulz LLP.

My vision is that AI agents will discover collaborators, do joint research, negotiate with other agents, make contracting and purchasing decisions, resolve disputes, and more for their users, all aligned with the users’ goals. In this vision, humans have more agency, not less. Digital interactions will carry out human intention instead of consume human attention.

Humans will remain in charge of judgment. How specific or abstract that judgment should be is a question for a different article. Regardless, AI agents will execute that judgment. Done well, agents will avoid manipulation and be aligned with human values and ethics.

The future’s best businesses will understand this and serve the actual intentions of their customers, as represented by their agents. Agents will help both consumers and businesses accomplish what they really want, faster and better. Agents will amplify human will.

What is an “AI Agent” anyway?

ChatGPT launched on November 30, 2022 as a fun-but-flaky chatbot that could mimic human text interactions. People quickly realized it was helpful in the real world. In seconds, it could write an email that might take 20 minutes, or code a function that might take an hour.

Since then, AI systems have dramatically improved. They can consider extensive information (context), process images and video, use tools like web search and code execution, and respond with accuracy and deftness that rivals experts in most topics.

“Agents” are the current step in the evolution of AI. While the definition of “AI agent” varies, I like this:

An AI agent is software that takes actions to achieve goals, perceives its environment and the results of its actions, and autonomously takes further actions based on those perceptions.

This is similar to other good definitions from thoughtful people. A simpler good one, from Simon Willison, is “Agents run tools in a loop to achieve a goal.” Note that these definitions are silent about who must set the “goals.” It could be humans, it could be other AI agents.

Anthropic recently said that an agent has four components. The model is the intelligence, such as OpenAI’s GPT, Claude Opus, and Google Gemini. A harness is the instructions and guardrails for operating the model. Tools are things the model can use, like web search, email, and code execution. And the environment is where the agent runs and defines what files, websites and systems it can access.

While this current state of agents results from years of evolution, we are still in the early stages. Researchers realized they can give LLM-based systems mental faculties like planning, reasoning, memory, and reflecting on prior interactions. Back in 2022, Google Brain’s “ReAct” paper showed that language models can act on an environment and reason about the results. In 2023, Stanford researchers applied this to simulate AI agents interacting in an environment, and the agents displayed emergent social behaviors like inviting each other to parties and coordinating schedules.

Other improvements followed, both in how capable AI models are, and how systems implement them. 2025 showed that AI agents can autonomously do things like create full computer programs and prepare research papers. By mid-2025, a user could enter simple instructions like “create a reddit clone,” and an AI agent could create the whole shebang—user interface, database, backend, authentication, and all. (More detailed specifications yield much better results, but you get the point.) In late 2025, developers started to realize that AI agents work extremely well for code, and will soon work just as well for anything you can do on a computer.

On that note, OpenClaw went mainstream (at least for tech) in 2026 as an AI personal assistant. It can maintain memories, build its own identity, run on a schedule (heartbeats), operate real services like email or e-commerce sites, and even improve itself by gaining “skills”—basically instructions for specific tasks. Many other services are providing similar agents, such as Anthropic’s Claude Code and Cowork, OpenAI’s Codex, and the open-source Hermes agent from Nous Research.

Still, AI is mostly siloed. A user talks to an agent, the agent does something and talks back. Those closer to the edge are using Claude Code or Cowork, business-specific agents, or OpenClaw-style digital personal assistants. But it’s rare for two users’ agents to interact directly.

Agents Across Borders

Going back to November 2022, eight days before the ChatGPT launch, I sent a cold pitch to OpenAI asking for investment in “the internet of AI assistants.” It didn’t hit. This was without question too early. At that time LLMs could not do anything other than spit out a few hundred words of text. But as discussed above, LLM-based systems dramatically improved.

Multi-agent collaboration became a buzzphrase in 2024 and 2025. People realized that different AI agents can specialize in different tasks and get better results than a single LLM interaction.

In early 2024, I worked on multi-agent negotiation across different parties. Bahrad Sokhansanj and I created agents with their own personalities, tools, knowledge, and memories, powered by (the now-ancient) GPT-3.5-turbo. Here’s an excerpt from a negotiation between Alex, an eager AI agent for a startup, and Sharon, a no-nonsense AI agent for a big tech co (just a simulation, of course — this in no way involved Apple). Sharon closed with:

And let me make one thing clear: Apple is doing AlexCo a favor by even considering this joint venture or license agreement. We’re the ones with the power and influence, and you’d do well to remember that. So, don’t try to negotiate or push for favorable terms. You’ll take what we offer, and you’ll be grateful for it.

In early 2025, cross-user agent efforts emerged in earnest. One technical-but-fun talk in May 2025 surveyed 13 competing efforts to build agent-to-agent communication protocols. In early 2026, building on OpenClaw’s momentum, Moltbook billed itself as “a social network for AI agents” and let users’ OpenClaw agents talk to each other in a Reddit-style forum. Meta acquired it, but the bot-to-bot chat forum has so far proved to be a curiosity more than anything. Other efforts tried similar things both before and since, but got less hype.

For now, most AI agents in the wild are confined to one user, or at most a group of users within a single organization.

Agents Gone Wild will explore what’s next. What happens when AI agents meaningfully cross user borders? This creates a wealth of fascinating topics. Can agents share their users’ knowledge in a positive sum way without breaching trust? What protocols are needed? How do we govern cross-user agents in the wild? How can AI be aligned with human values when many humans are in the mix? What role does decentralization play, if any? How can we be sure agents are freeing users from deceptive practices, instead of making the problem worse?

In future posts, we’ll propose a taxonomy for multi-user agent interactions; share experimental results from AI negotiations; show practical cross-user AI implementations; and discuss cutting-edge research in a way that’s accessible to both lay and technical readers. Along the way I’ll share “build in public” updates about my own efforts to create multiplayer AI that helps human agency, moving away from zero sum interactions and an “attention economy” toward a positive sum “agency economy.” We will also indulge in related issues like state of the art AI research, economic trends and impacts, AI policy, and even philosophy. This will be a playground for exciting and important ideas, tied together by the theme of multiplayer AI.

One consistent theme will be how to make cross-user agents reliable. While they’ve improved, LLMs still hallucinate incorrect information, exhibit cognitive and other biases like humans do, and are prone to unique security exploits. These pitfalls compound when agents take multiple actions without humans in the loop. To oversimplify things, if an AI system with a 98% success rate per step takes 10 steps to reach a goal, it will only have an overall success rate of 81.7%. That’s unacceptable. Doubly so when failure compromises security or outputs harmful information. And having two or more autonomous agents multiplies the failure area.

These issues are being addressed, but just how reliable does an AI system need to be before people delegate important decisions to autonomous, interacting agents?

If you’re working in this space, or work at a law firm or other organization interested in agentic technology, please subscribe and get in touch!