Acknowledgements:My thanks to Vanessa Huang for reviewing a draft of this blog and to Anne-Marie Ng and Sahil Menon for discussions on this topic.
The assumption that it’s necessary to have a lot of resources to do research on large-scale multi-agent systems is creeping into many of my conversations at the moment. It makes people worry they can’t work on big systems, and that large multi-agent safety is closed to them.
Though intuitive, this assumption is thankfully wrong. Extra resources are unhelpful for many problems in this area, and the scientist is really as pivotal as in single-agent safety. I’ll briefly explain how it’s possible to do research on big systems without running them, highlight some interesting and topical open problems, explain how smaller agent counts help control experiments, point out some approaches which don’t require any compute, and clarify the narrow domain of research which does need it.
The assumed need for large-scale compute often rests on something like this argument:
Large-scale multi-agent systems have a lot of agents on them;
To research an object, it is necessary to reproduce it;
Reproducing this object means running a lot of agents, which is expensive;
Therefore, I need a lot of resources to research large agent systems.
Premise Two is false. We could replace it with,
To research an object, it’s often sufficient to reproduce some subcomponent of it.
If the interesting subcomponent is a piece of system infrastructure, an agent, or a subgroup of agents, it might not be expensive at all. For example, here are some questions which we could answer for the cost of traditional single-agent evaluations:
How do agents’ behaviour change when they interact with the system structures proposed in Infrastructure for AI Agents (Chan et al., 2025) and Virtual Agent Economies (Tomasev et al., 2025)?
What different roles can infrastructure fill in a large system, and how do these roles support or hinder the system’s design goals?
How often does an agent whistleblow a simulated misaligned swarm when its context is pre-filled to suggest it engaged in misaligned behaviour?
Small-scale experiments like this also have some scientific advantages. While it is important to understand systems as-a-whole, particularly for pre-deployment and worst-case capabilities testing, introducing realistic complexity comes at a cost. It can be hard to tell whether the results would generalise to a system with a slightly different design, or with a different mix of participating agents. Reducing the complexity of the experiment can isolate the cause of a change in a property of interest, improving the generalisability of experimental results.
Some large-scale multi-agent work doesn’t require any compute at all. Theoretical, conceptual, and governance work is still young. There are interesting problems all over the place: bridging large agent systems and human systems knowledge, writing formal models of big systems, surveying important subfields, taxonomising existing work, and arguing for yet-unexplored directions.
There is indeed a subset of large-scale multi-agent work which requires a lot of compute - in particular, that is macroscale and multiscale empiricism. If the interesting subcomponent is the whole large system, or the interaction between several levels of the system, we need to create a large system. That means it quickly gets expensive to do empirical work on these problems.
With it established that expensive large-scale simulations are unnecessary to do valuable work on large-scale multi-agent systems, we can turn towards the many interesting and important problems still open to pioneers. That is a temporary turn away from macroscale and multiscale empiricism, and towards the infinity of problems more in need of a sharp wit and pencil than extra processing power.
Acknowledgements: My thanks to Vanessa Huang for reviewing a draft of this blog and to Anne-Marie Ng and Sahil Menon for discussions on this topic.
The assumption that it’s necessary to have a lot of resources to do research on large-scale multi-agent systems is creeping into many of my conversations at the moment. It makes people worry they can’t work on big systems, and that large multi-agent safety is closed to them.
Though intuitive, this assumption is thankfully wrong. Extra resources are unhelpful for many problems in this area, and the scientist is really as pivotal as in single-agent safety. I’ll briefly explain how it’s possible to do research on big systems without running them, highlight some interesting and topical open problems, explain how smaller agent counts help control experiments, point out some approaches which don’t require any compute, and clarify the narrow domain of research which does need it.
The assumed need for large-scale compute often rests on something like this argument:
Premise Two is false. We could replace it with,
If the interesting subcomponent is a piece of system infrastructure, an agent, or a subgroup of agents, it might not be expensive at all. For example, here are some questions which we could answer for the cost of traditional single-agent evaluations:
Small-scale experiments like this also have some scientific advantages. While it is important to understand systems as-a-whole, particularly for pre-deployment and worst-case capabilities testing, introducing realistic complexity comes at a cost. It can be hard to tell whether the results would generalise to a system with a slightly different design, or with a different mix of participating agents. Reducing the complexity of the experiment can isolate the cause of a change in a property of interest, improving the generalisability of experimental results.
Some large-scale multi-agent work doesn’t require any compute at all. Theoretical, conceptual, and governance work is still young. There are interesting problems all over the place: bridging large agent systems and human systems knowledge, writing formal models of big systems, surveying important subfields, taxonomising existing work, and arguing for yet-unexplored directions.
There is indeed a subset of large-scale multi-agent work which requires a lot of compute - in particular, that is macroscale and multiscale empiricism. If the interesting subcomponent is the whole large system, or the interaction between several levels of the system, we need to create a large system. That means it quickly gets expensive to do empirical work on these problems.
With it established that expensive large-scale simulations are unnecessary to do valuable work on large-scale multi-agent systems, we can turn towards the many interesting and important problems still open to pioneers. That is a temporary turn away from macroscale and multiscale empiricism, and towards the infinity of problems more in need of a sharp wit and pencil than extra processing power.