Multi-agent — when worth it

The multi-agent claim

The pitch: split a complex task across specialized agents (researcher, writer, critic, coder) — each focused on its strength. The reality: most multi-agent systems shipping in 2026 are just one agent with extra steps and worse latency.

The pattern earns its complexity when:

• Agents have genuinely different capabilities (e.g., one calls a code interpreter, one calls a search tool, one writes prose).
• The task has parallelizable subtasks (research three sub-topics in parallel).
• The role separation adds adversarial pressure (a critic agent that explicitly looks for weaknesses).