Single-Agent vs Multi-Agent: Decide and Design

PROJECT CONTEXT:
I am designing an agent system and every framework pushes crews, swarms, and role-play teams. The research record shows most production coding agents are single-loop, while multi-agent wins only for specific separable workloads.

GOAL:
Make an evidence-based single vs multi decision for my use case, and produce the corresponding architecture sketch.

REQUIREMENTS:
- A written workload analysis: separability, shared-state needs, latency and cost budgets
- A scored decision matrix comparing single-loop, sub-agent delegation, and full multi-agent
- The chosen architecture sketched with roles, communication paths, and failure handling
- Explicit criteria for when the decision would be revisited

CONSTRAINTS:
- Default to the simplest architecture that satisfies the workload
- Any multi-agent choice must name the coordination cost it accepts

DEFINITION OF DONE:
- The decision matrix scores all three options against my actual workload
- The chosen design handles the failure of any one agent or step
- Shared-state needs are mapped to a concrete mechanism, not 'they communicate'
- A revisit trigger is defined (scale, task mix, or quality threshold)

COMMON FAILURES TO AVOID:
- Choosing multi-agent because the demo looked impressive
- Underestimating coordination cost: message passing, state sync, partial failures
- Role-play theater: five agents doing what one prompt would do better
- No plan for what happens when one agent in a chain returns garbage