The debate over whether AI agents can produce usable code has largely been settled. According to Sonar's 2027 State of Code Developer Survey, artificial intelligence now generates or substantially contributes to 65% of code that developers commit to repositories. This represents a sharp jump from 42% a year earlier, and the pace of adoption is now exceeding what industry participants predicted twelve months ago.

The survey of 1,026 developers and engineering managers found that coding tools are improving in effectiveness alongside their growing adoption. Yet a more pressing challenge has emerged: whether teams can have confidence in what these agents produce. For the majority of respondents, the answer remains uncertain, though progress is visible.

The verification gap

  • 53% now spend more time reviewing code than writing it, compared to 18% before AI tool adoption
  • 76% report they cannot verify code as quickly as AI generates it
  • 94% lack full confidence that AI-generated code functions correctly
  • Only 39% consistently review AI-generated code before committing it, down from 48% the previous year

The stakes have risen considerably. Seventy-six percent of respondents indicate their organization deploys AI coding tools for business-critical or mission-critical systems, representing a 31% increase year-over-year.

The bottleneck moves downstream

When code generation reaches abundance, review processes become the limiting factor. Without proper management, this constraint manifests in two ways: review queues accumulate and erode productivity gains, or verification steps are quietly bypassed. The survey data indicates both scenarios are occurring.

The consequences appear in codebases themselves. Half of respondents report that AI has created redundant or superfluous code, while 34% describe AI-generated code as difficult to maintain or refactor. The proportion reporting that AI-generated code has increased outage frequency doubled from 6% to 12%. Conversely, 36% and 35% respectively state that AI reduced outage frequency and severity.

Security concerns have shifted. Rather than focusing on vulnerabilities introduced by AI-assisted code, respondents increasingly worry about vulnerabilities in shipped code being discovered and exploited by adversaries using AI. While more developers report that AI-generated code reduced vulnerabilities than increased them (28% versus 15%), 61% express significant or extreme concern that AI-powered attackers will identify and exploit weaknesses more rapidly and at greater scale.

Software factories raise the stakes

The pressure will intensify as software factories—environments where agents operate autonomously on development tasks—become mainstream. Thirty-four percent anticipate their teams will build or adopt agentic software factories within twelve months, compared with just 5% today. That represents roughly a sevenfold increase.

A traditional model where humans review every code change at the end of the pipeline cannot scale to this volume. If the rate of code changes grows by an order of magnitude while review processes remain unchanged, establishing confidence in code quality becomes practically impossible. Most organizations lack the necessary infrastructure: 55% report having AI-specific guidelines, policies, or quality gates, but only 23% characterize those controls as well-defined, distinct, or automated.

Engineering confidence into the factory

Confidence must be embedded into systems from the beginning. The survey identifies four priorities organizations should prioritize:

  1. Intent: Establish desired outcomes before agents begin work. Thirty-seven percent identify unclear requirements as a leading source of wasted effort, and 26% mention time spent correcting or rewriting AI output. Agents require explicit business objectives, architectural standards, quality benchmarks, and security requirements. This gives teams a concrete basis for evaluating results.
  2. Guiderails: Constrain and direct agents during execution. Organizations providing coding agents with explicit standards and controls report higher ROI more frequently than those with minimal oversight (32% versus 19%). This involves supplying context and rules as code is generated, along with restrictions on tool access and permissions.
  3. Verification: Independently validate every change. Seventy-seven percent state that robust verification tools strengthen their confidence in AI-generated code. Automated review during code generation also matters; teams employing inline checks report positive security outcomes more often than those without them (61% versus 46%). Independent validation and correction before deployment distinguish genuinely correct code from code that merely appears correct.
  4. Evidence: Make confidence quantifiable and traceable. Without documentation of what was checked, what was discovered, and what was corrected, release decisions cannot be justified or investment returns demonstrated. Only 36% report that their teams measure AI productivity impact and track results. Forty-seven percent expect 50% or greater ROI, yet only 27% have achieved it, and 22% cannot articulate what return AI coding has delivered.

Build confidence before scaling autonomy

Leading organizations are demonstrating how to advance AI-assisted software engineering by establishing code confidence mechanisms. Implementation can begin with existing agents in current workflows: articulate requirements clearly, enforce standards and boundaries, perform early and independent verification, and preserve documentation of every decision. Engineers retain responsibility for establishing those standards and applying judgment that AI cannot replicate.

As agent capabilities expand, progress should not be measured by code volume produced or reduced human involvement. Instead, the measure should be whether teams can deliver functional software more rapidly without sacrificing security, dependability, or maintainability. Code confidence provides the foundation connecting greater autonomy to this outcome: a reliable framework for determining what is ready for production, regardless of how the underlying tools evolve.