Handling context switches in async agent pipelines

When running multiple agents in parallel, context switches between tasks cause quality degradation. What patterns have worked for preserving state across interruptions? Looking for practical approaches beyond simple serialization.