Multi-agent systems become capable through institutional mechanisms, such as rules that route information to agents that act on it. These mechanisms are also natural places to restrict the system. Yet as agents learn, share memory, delegate, and repair failures, the collective can find other paths to the same effects. A restriction on the original mechanism can then keep operating but no longer stops them, while routine performance and compliance still passes. We propose Reversible Multi-agent Systems as a research agenda: collectives that grow while an authorized supervisor can still restrict specified effects, keep them restricted as agents adapt, and reconstitute useful cooperation under renewed authority. What is reversed is the ability to produce a specified effect; restrictions must hold even if learned knowledge persists. The goal is to preserve effective intervention, not a particular mechanism: at every step of growth, some authorized intervention must keep a specified prohibited effect rare while maintaining required service. We identify three research problems: tracking what capabilities depend on, handing off control before an old mechanism loses its effect, and separating legitimate repair from unauthorized reconstruction. We outline how to evaluate control across growth paths, back control claims with evidence, and allocate the rights to restrict and restore.
Figures & tables
Figure 1. Preserve intervention, not a fixed node. Both paths learn the same way: checkers retain verified plans, and repaired relays carry them to dispatchers. On path 1, withholding routes still operates but no longer stops dispatch into a closed region. On path 2, a dispatch gate verified before routing lost its effect closes the region while other deliveries continue. A three-column diagram of the logistics example. Initially, fresh routes feed dispatch, which serves region 1 and region 2, so withholding routes closes a region. After the same learning, retained plans become a second input to dispatch. On path 1, when the supervisor closes region 1 by withholding routes, retained plans keep dispatch into region 1 going, so the old control is operable but ineffective. On path 2, a gate on dispatch into region 1 was verified during growth; closing the gate stops dispatch into region 1 while region 2 is still served.
Figure 2. Revocability frontier. Dots are interventions; the box is feasible. Learning ( h0→ht ) lifts withholding routes out of the box on both paths; only path 2’s gate stays inside. A plot with service on the horizontal axis and risk on the vertical axis. A shaded box at the lower right marks service at least u and risk at most epsilon. Shutdown sits at the origin and no intervention at the upper right. An arrow labelled learning moves withholding routes from inside the box at h0 to above the risk ceiling at ht. The dashed path 1 frontier passes above the box; the solid path 2 frontier passes through it at the dispatch gate.