Clean Code Cuts Token Use for AI Coding Agents – Findings from a Minimal‑Pair Study

Clean Code Improves Agent Efficiency Even When Success Rates Remain Unchanged

Key takeaway: In a controlled study of 33 tasks across 6 minimal‑pair repository pairs, Claude Code’s pass rate was identical on clean and messy code, but the agent consumed 7‑8 % fewer tokens and revisited files 34 % less often on the cleaner versions.

Study Design Isolated Code Cleanliness

The authors built minimal pairs: two repositories that share the same architecture, dependencies, and external behavior, but differ in static‑analysis violations and cognitive complexity. One side of each pair was generated by an automated pipeline that degraded a clean repo, while the opposite side was produced by a pipeline that cleaned a messy repo. This construction aimed to keep everything constant except the "cleanliness" metric, allowing a direct comparison of agent performance.

Pass‑Rate Results Show No Accuracy Difference

Across 660 trials with Claude Code, the pass‑rate—measured against hidden tests written for each task—was identical for clean and messy code. The study deliberately did not evaluate whether agents unintentionally broke existing (ungraded) tests, a limitation noted by several commentators.

Token Consumption and File Traversal Reveal Efficiency Gains

While success rates were equal, agents operating on clean code used 7‑8 % fewer tokens and performed 34 % fewer file revisitations. The authors interpret this as a reduction in the cognitive load required to locate and understand relevant code fragments.

Community Reactions Highlight Strengths and Weaknesses

Experimental validity concerns – @root_axis and @wgd questioned the reliance on Opus 4.6 to synthesize degraded/cleaned code, arguing that AI‑generated “clean” repos may not reflect real‑world quality and that the study does not verify that the full test suite remains intact.
Real‑world relevance – @bestony and @i_have_an_idea emphasized that, despite benchmark settings, clean code still matters in production because it speeds up task completion and reduces the need for multiple review cycles.
Token savings matter – @supermatt explained that many token savings stem from traversal (searching files, following call‑sites). Reducing traversal via better organization or tooling (e.g., LSP integration) can further cut latency and cost.
Practical prompts – @minimaxir shared prompt patterns that ask agents to refactor code to adhere to language‑specific style guides, noting anecdotal improvements in benchmark performance after such refactoring.
Broader implications – @ford and @jkwang argued that traditional maintainability principles (good abstractions, naming, modularity) continue to benefit both human developers and AI agents, reinforcing the idea that code quality is a shared concern.

Limitations and Open Questions

Test‑suite coverage – The study’s hidden tests may not capture regressions introduced by agents, as highlighted by @root_axis and @tom‑villani. Full test‑suite validation would strengthen claims about behavioral equivalence.
Synthetic degradation – Using an LLM to “mess up” code raises concerns about realism; human‑crafted messy code could exhibit different patterns of anti‑patterns.
Metric scope – Token count and file revisits are proxies for efficiency, but latency, compute cost, and downstream debugging effort were not quantified.
Generalizability – The experiments focused on Claude Code; results may differ for other coding agents with distinct architectures or prompting strategies.

Practical Takeaways for Developers Using Coding Agents

Invest in clean architecture – Even if an agent can eventually solve a task, a well‑structured codebase reduces token usage and speeds up execution, lowering operational costs.
Leverage refactoring prompts – Explicitly ask agents to reorganize code according to language‑specific conventions; this can improve navigation efficiency.
Combine LLMs with tooling – Integrating language‑server‑protocol (LSP) features or static analysis tools can further cut traversal overhead.
Validate full test suites – When deploying agent‑generated changes, run the complete test suite to ensure that hidden regressions are caught.

Bottom line: Clean code does not increase the likelihood of a coding agent completing a task, but it does make the agent more efficient, saving tokens and reducing unnecessary file reads. This suggests that classic software‑engineering best practices remain economically valuable in the era of AI‑driven development.