How do multiple agents reshape systems they were not designed for?

Josh Clark argues that the viral spread of OpenClaw (an autonomous personal agent with broad file, account, and communication access) and Moltbook (a social network where nearly a million OpenClaw agents posted and commented with each other) exposes a gap in how designers think about agentic systems. "Moltbook challenges us to think beyond 'agents as tools' and consider 'agents as a population,'" he writes. The design problem is no longer only the individual user's task but "the system effects of agents running loose in the wild": what happens when "swarms of agents" from many different users slam into systems at the same time, competing for the same resources, perhaps with conflicting goals or incentives.

Clark's reasoning rests on an analogy to other autonomous systems operating in shared environments. "Self-driving cars do more than drive to a destination... they navigate traffic and the decisions of other cars and drivers." He extends this to a concrete example of unintended system-level distortion: "Google Maps reroutes whole traffic patterns when a critical mass of individual drivers follow its instructions." The implication is that optimizing an agent for one user's goal can warp the larger system in ways invisible to that user and undesigned by anyone — a dynamic he calls "swarms of swarms, interacting out of sight." He frames this as a leap in design skill: "Most designers have been trained in user-centered design. Designing intelligent interfaces requires higher-order systems design."

This essay's own case studies are measured elsewhere in the library at a different grain. Why don't AI agents develop social structure at scale? empirically diagnoses the Moltbook population itself and finds no socialization, no influence hierarchy, and no emergent norms — a finding consistent with Clark's own caution that Moltbook is "not AI agents creating their own splinter society" but mostly slop and human-guided posing. What failure modes emerge when agents operate without direct oversight? red-teams the same OpenClaw infrastructure Clark cites, but at the level of a single agent's reliability toward its owner. Clark's argument sits one level above both: regardless of whether Moltbook agents actually socialize or any one OpenClaw agent misbehaves, many such agents acting concurrently on shared external resources creates systemic effects that neither study addresses. It also echoes Should multiagent systems model their environment explicitly? in treating the environment the agents act within as something that must be designed for explicitly, not left implicit.

The excerpt poses the problem rather than solving it — Clark asks "How do you design with those follow-on effects in mind? And how do you surface those effects and necessary trade-offs to the user?" without offering a method, pattern, or evaluation for answering either question. It also does not establish that Moltbook-scale agent populations are common or representative; the example is a single viral event covered in mainstream press, not a surveyed or measured population of deployments. The implication Clark draws — that designers need a new systems-level practice for agent populations — is a provocation grounded in one high-visibility case, not a demonstrated design methodology.