Multi-Agent Interaction Without Shared Context

Each agent is an independent entity with its own context and state; interaction only through explicit structured mechanisms. Benefits of process isolation: independent development and testing, failure does not propagate, true parallelism. Infrastructure: shared filesystem (data plane) and communication/control mechanism (control plane: message passing, state query, interrupt, resource scheduling). OS correspondence: program – static prefix, process memory – trajectory, CPU – LLM in time-sharing mode, fork – spawn_subagent. The only thing that doesn’t translate: processes pass bytes exactly, agents pass semantics, and each retelling can distort.

Related: [Inter-Agent Communication], [Filesystem in Multi-Agent Systems], [Failure Modes of Multi-Agent Systems]