Running More Than 2 Parallel Sub-Agents Is Likely Just Burning Tokens

Starting September 16, 2024, developers debated whether running multiple sub-agents in parallel is worthwhile. The core claim came from pvncher (described by some as an Anthropic/Claude engineer): running more than 2 sub-agents concurrently is almost certainly paying for tokens with no quality gain. ALEXEIMARTOV, StevenMathern, kevinnbass and patricksrail echoed the view, and HankYeomans confirmed after review that the judgment holds.

Confirmed

• Core claim: more than 2 parallel sub-agents is essentially token waste with zero quality gain
• Reason: agents lack trust in each other and repeatedly re-check one another's work, creating redundant verification calls
• Applicable scenario: multi-agent parallelization such as running 100 similar tasks across 3 parallel channels
• HankYeomans corroborated the judgment after his own retrospective
• omarsar0, drawing on agent harness building, added that sub-agents are best used for parallel research, code review, monitoring/tracing and better context management, with "1 orchestrator + 1 executor" working best

Why it matters

• The view challenges the popular belief that more parallel agents are inherently better; kevinnbass argued that mindset belongs to an outdated multi-agent framework era
• omarsar0's practice offers an actionable alternative: sub-agents are useful in the right scenarios rather than blindly scaling parallelism
• Direct cost implications for teams building multi-agent workflows: more parallel agents may only inflate token spend
• pvncher's first-hand experience and engineer credentials add authority to the judgment

2026-09-16 ~ 2026-09-17 · 7 related posts

• [source] Running More Than 2 Sub-Agents in Parallel Burns Tokens for Zero Quality Gain — pvncher · 2026-09-16
• Running more than 2 sub-agents at once burns tokens for zero quality gain, dev argues — kevinnbass · 2026-09-16
• Running 3+ subagents likely burns tokens for zero quality gain, argues dev — StevenMathern · 2026-09-16
• Too many sub-agents burn tokens for zero quality gain, warns Anthropic engineer — patricksrail · 2026-09-16
• Running more than 2 sub agents wastes tokens for zero quality gain, dev argues — HankYeomans · 2026-09-16
• More than 2 parallel subagents just burns tokens, harness builders say 1 orchestrator + 1 executor works best — omarsar0 · 2026-09-17

1 near-duplicate retellings: ALEXEIMARTOV