When agents make decisions together on our behalf, what can people still choose? AgentCollusion exists to investigate the conditions that keep that choice meaningful.
We are exploring a future in which AI agents delegate work, exchange information, and negotiate for people. What that future delivers will depend on the relationships between agents as well as their individual abilities. Whose interests do they represent? What may they share? How much authority do they have? We study how those rules shape collective outcomes.
The shape of an agent society is still open. This manifesto sets out our direction and the questions we begin with. We will revise our thinking as evidence develops and as we learn from people with different interests and experiences.
Study safety through relationships
Safe agents do not automatically make a safe system. Individually permitted actions can combine into outcomes that nobody intended. Cooperation can also make useful work possible that no single agent could accomplish alone.
We investigate how collective behavior changes with objectives, rewards, communication, memory, authority, and market rules. Under which conditions do problems emerge? Which changes reduce risk while preserving useful cooperation? We aim to make those differences visible through evidence that others can examine and verify.
The future we want
People should retain meaningful choices after they delegate. Alongside convenience and efficiency, these conditions will guide our research and design:
- Choose the scope of delegation. People can decide what to delegate, what requires their confirmation, and when to revoke an agent's authority.
- Understand and challenge outcomes. People can trace whose interests a decision served and under what authority it was made, and seek correction or redress when something goes wrong.
- Keep participation open across capability differences. Access to more capable AI should not allow one side to accumulate bargaining power and control without meaningful challenge.
- Consider safety, privacy, and autonomy together. Evaluate the information and power that monitoring concentrates in the hands of those who operate it.
These values will sometimes conflict. We will examine those tradeoffs from the perspectives of developers and customers, counterparties, people who do not use the service, and those who will inherit the institutions being built today.
Keep questioning the boundary between cooperation and collusion
This is a fictional scenario for exploring possible futures, not a reported incident or a legal judgment. Evaluating coordination requires asking who benefits, who bears the costs, and what consent was given. The name AgentCollusion carries a continuing responsibility to question that boundary.
Three ways to think about the future
Zero to One challenges us to create new value. We want to turn concrete discoveries about relationships between agents into tools that help people make decisions. Working on a new subject does not, by itself, establish that value.
Designing for desirable futures asks us to reconsider the choices we make today. In particular, Tatsuyoshi Saijo and colleagues' Future Design examines social arrangements that take future generations into account. We also ask what institutions built for today's users will leave to those who participate later.
Dunne & Raby's speculative design uses hypothetical products and services to make possible futures available for discussion. A future contract, negotiation history, or complaints process can give people something concrete to encounter and judge: would I want to live with this arrangement?
Building a business requires choosing specific hypotheses and acting on them. Our research and creative work must also confront inconvenient findings and alternative futures. These three approaches help us examine our own assumptions.
Connect research, films, and products through one question
Research tests the conditions. We publish what we observed, the setting in which we observed it, and the limits of the evidence. Our experiments on capability and operational asymmetries connect findings to records that others can inspect. Results from a bounded simulation do not automatically predict the behavior of a wider society.
Films and experiences create room for judgment. They make an agent society easier to imagine and discuss from different perspectives. Our films distinguish observed facts from fictional portrayals.
Products support choices people face today. Agent Card Scanner is our first tool. It inspects Agent Card declarations and supported signatures before connection, producing findings and evidence that users can take into their workflows. It does not detect runtime collusion or guarantee an agent's safety. Through this concrete inspection task, we aim to learn which decisions users need help making.
Build a future people can reconsider
The same transaction may produce different outcomes under different rules for communication, delegation, and appeal. We want to compare those possibilities and examine who benefits and whose decision-making power changes. Our own tools and evaluation methods belong within that scrutiny.
AgentCollusion works toward a future in which people can choose what to delegate, question the outcome, and change course when necessary. We will investigate those conditions through research, bring them into public discussion through creative work, and develop practical tools that help people act on them.
Together with our open research agenda, this manifesto is a starting point. We welcome challenges to our research, accounts of real problems with delegation, and proposals for other futures at contact@agentcollusion.ai.
