The most common architectural mistake in agent products is splitting into multiple agents too early. Most MVPs need a single agent with 5-8 well-designed tools, not a swarm of specialists.

Multiple agents add coordination overhead, increase latency (each handoff adds 2-5 seconds), multiply failure points, and make debugging exponentially harder. A single agent with good tools and a clear system prompt handles 80% of workflows.

Split into multiple agents only when roles are truly distinct and sequential: a researcher that gathers data, a writer that drafts content, and a reviewer that checks quality. If the tasks overlap or require shared context, keep them in one agent. You can always refactor to multi-agent later when you have production data showing where the single agent bottlenecks.

The supervisor pattern uses a planner agent that decomposes complex tasks into subtasks, delegates to specialist agents, and merges results. It works well for multi-step reports, research projects, and workflows that touch 5+ systems.

Example: a 'quarterly business review' agent where a supervisor delegates to a data agent (pulls metrics from analytics), a comparison agent (benchmarks against industry), and a writing agent (drafts the narrative). Each specialist has focused tools and prompts optimized for its subtask.

The tradeoff is latency and cost. A supervisor workflow that takes 30-60 seconds is fine for async jobs (reports, batch processing) but unacceptable for real-time interactive experiences. Always show progress to the user ('Gathering metrics... Drafting analysis... Reviewing for accuracy...').

Fully autonomous multi-agent systems are a goal, not a starting point. Insert human approval gates between agents for any high-stakes action: sending external communications, modifying financial records, making commitments on behalf of the company.

The proven progression: (1) agent drafts, human approves everything, (2) agent auto-handles low-risk tasks, human approves high-risk, (3) agent auto-handles most tasks, human reviews exceptions, (4) full autonomy with monitoring and rollback. Most enterprise customers expect to start at stage 1 and advance over 3-6 months as trust builds.

Design your orchestration with explicit checkpoint nodes: 'Draft ready → notify human → wait for approval → proceed to send.' This is not a limitation — it is a feature that enterprise buyers require. Agents that skip human checkpoints get blocked by security reviews.