Parallel Agents Backfire on Coding Tasks

The performance of parallel agent systems varies drastically depending on the nature of the task. Parallel processing is advantageous for read-centric independent tasks (such as research, search, and file review), but a single agent yields better results for tasks requiring sequential decision-making (such as coding and writing).

Strengths and Limitations of Parallel Agents

According to research by Anthropic, multi-agent systems demonstrated a 90.2% performance improvement over single agents in internal research evaluations. However, this comes at the cost of consuming approximately 15x more tokens compared to standard chat, and Anthropic acknowledges that coding tasks are harder to parallelize than research. LangChain also advises that while read tasks can be parallelized, write tasks cannot.

Performance Degradation in Sequential Tasks

A peer-reviewed paper published in Nature Machine Intelligence demonstrated through 260 controlled configurations that multi-agent variants worsened coding results (SWE-bench Verified) by -1.3% to -12.8%. Once single-agent performance exceeds approximately 45%, the advantages of multi-agent systems disappear or even become negative. Additionally, without centralized verification, error amplification can reach up to 17.2x.

Analysis of System Failure Causes

The MAST study by UC Berkeley recorded high failure rates of 41–86.7% across seven popular multi-agent frameworks. These failures stem from design and coordination flaws rather than model limitations, and cannot be resolved by standard protocols alone.