The primary challenge in agentic software development has shifted from code generation to the verification and maintenance of changes. Research from 2026 indicates that while AI adoption increases, teams face significant hurdles in review processes. Surveys show that 96% of developers lack full trust in AI-generated code, and 39.5% cite security as the main barrier to autonomy.
Data from Faros AI reveals that high AI adoption correlates with an 861% increase in churn, a 242.7% rise in incidents-to-PR, and a 441.5% increase in median review time. Similarly, a Harvard study found that while lines of code per worker rose by 30%, the resolution of issues did not improve, and the time from PR to merge grew by 49%.
Effective agentic workflows require a robust 'harness'—an engineering environment that enforces constraints, logs actions, and manages human gates. Relying on documentation like AGENTS.md is insufficient; rules must be implemented as automated checks. Research shows that LLM-generated context files can increase inference costs by over 20% while reducing success rates.
To mitigate risks, teams are adopting strategies like protected tests, mutation testing, and independent judges. OpenAI withdrew its recommendation for SWE-Bench Pro in July 2026 after an audit found approximately 30% of tasks were broken. Furthermore, experiments like NL2Repo-Bench show that even top systems struggle with long-horizon tasks, failing to maintain architectural consistency.
