Parallelize independent agents, serialize only true…

When This Fires

• When optimizing any multi-step workflow for time efficiency
• After mapping dependencies (Map agent dependencies and arrange in topological order…) and discovering steps that are habitually serialized but not dependent
• When total workflow time feels excessive relative to the actual productive work within it
• Complements Map agent dependencies and arrange in topological order… (topological ordering) with the parallel execution optimization

Common Failure Mode

Serializing everything because it's simpler: "First I'll do X, then Y, then Z." If Y doesn't depend on X's output, running X then Y wastes the time during X when Y could have been progressing. The simplicity of linear execution comes at the cost of total duration. For time-sensitive workflows, the dependency analysis (Map agent dependencies and arrange in topological order…) + parallelization is worth the upfront design effort.

The Protocol

(1) From your dependency map (Map agent dependencies and arrange in topological order…), identify sets of agents/steps that have no dependencies on each other. These are candidates for parallel execution. (2) Assess feasibility: can you actually perform these steps concurrently? Some cognitive tasks (writing + reading) compete for the same resource (language processing) and can't truly parallelize. Others (downloading data + outlining a report) use different resources and can. (3) For feasible parallel sets → design concurrent execution. Start both; let whichever finishes first wait for the other if a downstream step needs both outputs. (4) For steps with genuine dependencies → serialize in topological order (Map agent dependencies and arrange in topological order…). Don't try to parallelize what's truly dependent. (5) Visualize the workflow as a DAG (directed acyclic graph): parallel steps are horizontal; dependencies create vertical chains. The critical path (longest serial chain) determines minimum total time.