Single Agent vs. Multi-Agent Workflows
retained text
A few months ago, my approach was to use Opus as an orchestrator that managed a backlog and spawned different agents depending on the task. Haiku for small documentation changes, Sonnet for regular coding, Opus for more complex tasks, and Fable for design decisions and bigger-picture thinking. This worked really well, and I got a lot done. However, I've recently read that this approach might not be as efficient anymore because every new agent starts with a cold context/cache, and it might be better to have one persistent agent handle most of the work. What's everyone's experience with this nowadays? Is a single agent with selective delegation the better approach, or is the orchestrator/multi-agent workflow still worth it? I'm particularly interested in balancing productivity, context caching, and token costs without sacrificing quality.
document fetched 2026-11-12 00:42 · canonical https://reddit.com/r/ClaudeCode/comments/1xto6y3/single_agent_vs_multiagent_workflows