AI agent infrastructure company Composio ran DeepSeek V4 Flash through eight different harnesses, using the same 30 agent tasks across real apps like Gmail, GitHub, Slack, Calendar, and Notion. In agent development, the harness is the execution framework wrapped around the model — the layer that manages context, tool calls, and task execution.
The results: Pi Agent delivered the highest success rate and the lowest cost. Claude Code was the fastest of the bunch, but it also carried the heftiest price tag.
In short, Pi Agent was both the most reliable and the most economical option, while Claude Code was the quickest and most expensive.