Can AI coding agents actually fix bugs?
Yes, and they are surprisingly good at diagnosing what's wrong. In a benchmark of 5G telecom network engineering tasks, four different AI models diagnosed bugs at rates exceeding 91% [3]. That means the agents could correctly identify where the problem was in the code the vast majority of the time. However, actually fixing the bug is a different story: the same study found that resolve rates — the percentage of bugs the agents could successfully patch — ranged from only 10% to 30% [3]. The gap between diagnosis and resolution suggests that AI agents are strong at pattern recognition but weaker at the iterative editing and domain-specific reasoning needed to produce a correct, safe fix.
The implication is clear: AI agents can help find bugs, but relying on them to fix every bug autonomously will leave many unresolved. The study also found that injecting domain knowledge (like 3GPP specification excerpts) improved resolve rates for bugs tied to those specifications, but had little effect on generic defensive checks [3]. So the type of bug matters a lot.
Can we have the speed without the debt?
Partially, yes — but it requires deliberate guardrails. The Wily study shows that giving AI agents real-time feedback on code complexity (cyclomatic complexity, maintainability index, and logical lines of code) meaningfully reduces the complexity growth of patches, with only a tiny drop in resolution rates (less than 1.3 percentage points) [2]. This is a practical technique: treat complexity as a quality gate, just like test passing. Organizations can also establish clear guidelines for when and how AI-generated code is deployed, prioritize technical debt management, and train developers to review AI output critically [1].
However, there is a catch. AI tools themselves come with challenges: they are heavily dependent on the quality and quantity of training data, and over-reliance on them can lead to skill degradation and monotonous codebases that produce mediocre products [5]. The same paper warns that as organizations rely more on AI, they may end up with code that is uniform but not excellent — a different flavor of technical debt. The bottom line: AI coding agents can reduce bugs and boost speed, but without active complexity monitoring and human oversight, they will increase hidden technical debt. The evidence consistently shows that the fastest path is not the safest one.
About These Sources
This answer is built on 5 studies (4 peer-reviewed, 1 preprint) — published from 2025 to 2026, 5 from 2024 or later, 1 in Q1–Q2 journals — selected as the most relevant from 5 studies that passed quality screening, drawn from 45 papers retrieved from a database of over 500 million.
Sources used in this answer
The Hidden Costs of Coding With Generative AI
Generative AI tools can boost developer productivity by up to 55%, but rapid deployment creates dangerous technical debt, especially in brownfield environments with legacy systems; the paper recommends clear guidelines and training to avoid costly failures.
Wily: High-Performance Complexity Gated-Feedback for AI Coding Agents
In a controlled study of 500 SWE-bench tasks with two frontier models, AI agents with complexity-gated feedback reduced mean complexity growth by 10–27% and logical lines of code growth by 8–23% versus test-only feedback, but both AI conditions remained substantially above the human baseline, confirming a persistent complexity gap.
SWE-Bench 5G: Benchmarking AI Coding Agents on Telecom Network Engineering Tasks
On a new benchmark of 5G telecom tasks, four AI models diagnosed bugs at rates exceeding 91% but resolved only 10–30% of them; injecting domain-specific specification excerpts improved resolve rates on specification-dependent bugs but not on generic defensive checks.
Faster code, deeper debt? a multivocal literature review on technical debt and its early signs in llm-assisted software development
A multivocal literature review on LLM-assisted software development found that adoption of AI agents leads to a statistically significant increase in bugs surfacing after deployment, indicating hidden technical debt.
LEVERAGING AI TO REDUCE TECHNICAL DEBT
AI-powered tools can proactively detect issues and suggest refactoring to reduce technical debt, but they depend heavily on training data quality and may cause skill degradation and monotonous codebases if over-relied upon.