Agentic Coding, Developer Exploration, and Software Security

Agentic coding has shifted artificial intelligence (AI) from an assistant augmenting local changes to delegated agents executing tasks end-to-end. We argue that agentic coding collapses the cost of exploration but not the judgment that sustains it: delegated construction becomes cheaper, but ongoing maintenance may require judgment agents do not supply and novice developers, whose exploration expands most, least possess. Leveraging Claude Code’s launch and parsing its commit-message trailers in public commit histories, we observe 61,056 treated developers before and after treatment and compare them with 61,056 controls matched on pre-treatment activity, using monthly activity from January 2025 through March 2026.

Combining rolling-entry matching with matrix-completion counterfactual estimation, we find agentic coding increases activity: repository initiation rises by 7.4% and push activity by 93.1%. The increase is concentrated in exploration: pushes to original repositories rise by 72.6%, versus 4.5% for pushes to forks, with sharper contrasts for novices. We follow repositories developers create to test whether exploratory expansion compromises security.

Using OpenSSF Scorecard measures, we separate security practice (construction-bundled setup including security policies or static analysis) from security state, a repository’s absence of known code and dependency vulnerabilities, which ongoing maintenance sustains. Repositories show modest practice gains, similar across experience levels, but sizable state declines, steepest among novices. The findings shift evidence on AI coding from task-level productivity to the extensive margin of software production.

Agentic coding broadens exploration, especially among novices, while underproviding the sustained judgment needed to keep artifacts secure: the margin on which those developers are most exposed.