Analyzing and Mitigating Cost-Inefficient Behaviors in Coding Agents

What happened

This marks the first research focused on cost-inefficient behaviors of coding agents. The research team analyzed 1,200 run traces from four configurations of Claude Code and Mini-SWE-Agent on the SWE-bench Verified benchmark, identified three categories of cost-inefficient behaviors, then evaluated three mitigation strategies against 10,000 held-out task traces from SWE-bench Verified and SWE-bench Pro. The study found these three inefficient behavior categories cover 79% to 98% of coding tasks, and can account for up to 22.75% of total task costs.

Key facts

Study Type

First systematic research on cost-inefficient behaviors in coding agents

Analysis Sample

1,200 run traces from Claude Code and Mini-SWE-Agent

Inefficiency Cost Share

Inefficient behaviors account for up to 22.75% of total task costs

Top-Performing Mitigation Strategy

Advanced skills designed by developers

Best Strategy Cost Reduction

Delivers up to 41.73% cost savings

Strategy Validation Scale

Validated against 10,000 held-out SWE-bench task traces

Background

While today's coding agents deliver strong task performance, they often incur steep monetary costs to run. The frequent, cost-inefficient behaviors that drive these costs had not been systematically explored or documented prior to this study, leaving teams without targeted guidance for optimization.

Why it matters

For the broader industry, this study is the first to identify the core sources of cost waste in coding agents, laying a clear roadmap for future cost-reduction optimizations. For developers, the research confirms that developer-customized advanced skill prompts deliver twice the cost-reduction efficiency of model-generated skills, and warns that blindly applying structure-aware retrieval can actually drive up costs. For end users, these findings point to further reductions in coding agent operating costs, lowering the barrier to access AI programming tools.