Source: Sonar, 2027 State of Code Developer Survey report, released 2026-11-12; survey of 1,026 developers and engineering managers. Press release; report. Highlighted on LinkedIn by Cole Medin, 2026-11-16.
Read with care: this is a vendor survey from a company that sells code verification, so the framing favours verification tooling. The numbers are still a useful snapshot of how teams say they work.
The headline numbers
65% of committed code is, on average, generated or significantly assisted by AI.
76% say they generate AI code faster than they can verify it.
53% now spend more time reviewing code than writing it, against 18% before adopting AI tools.
Only 39% completely agree that AI code is always checked before commit (Cole Medin's post puts the prior-year figure at 48%), although 94% don't fully trust it.
42% are very or extremely concerned that well-written AI code makes bugs harder to spot, yet only 23% have AI-specific guidelines or automated checks.
76% use AI coding tools on business- or mission-critical services.
5% run agentic "software factories" today; 34% expect to within a year.
Only 36% measure AI's productivity impact and can see the results; 22% don't know what return they got.
The framework: four multiplying terms
Sonar frames code confidence as intent, guiderails, verification and evidence multiplied together, so a zero in any one term zeroes the result.
Intent: define outcomes, architecture standards, quality and security criteria before agents start. Unclear requirements (37%) and rewriting AI code (26%) top the toil list.
Guiderails: give agents standards, context and tool or permission limits while they work. Teams with clear controls report higher ROI more often.
Verification: check every change independently of the agent that wrote it; 77% say strong verification tools raise their trust.
Evidence: record what was checked, found and fixed so release decisions and AI's value can be audited.
The practitioner rule Cole Medin draws from it: agents may write anything, but nothing merges until a check the agent doesn't control passes. A model approving its own work doesn't count.
What it means for engineering leaders
Review is now the bottleneck, and a skim no longer catches the bugs that matter, because plausible-looking code is the norm.
The guardrail gap is large: concern is high, but fewer than a quarter have rules or automated checks specific to AI code.
Factories are coming before the controls: a sevenfold expected jump in agentic factories, with checking habits slipping year on year.
Measurement is weak: most teams cannot show what AI coding returned, which makes cost and risk decisions guesswork.
Tradeoffs and caveats
Self-reported survey data; "always checked" and "faster than we can verify" are perceptions, not measured rates.
Vendor-run, so the multiplication framing conveniently makes verification non-optional.
Stronger automated gates add CI cost and latency, and can push agents to optimise for passing the gate rather than doing the right thing.
Failure modes this data points to
Agents marking their own work as reviewed or tested.
Static analysis treated as sufficient verification for behavioural changes.
Intent left implicit, so verification checks code against nothing.
Evidence never stored, so a later incident can't show what was checked.
Review checklist for an agentic delivery pipeline
Is there a merge gate the authoring agent cannot run, skip or edit?
Are intent and acceptance criteria written before the agent starts, and checked against at review?
Are there AI-specific rules (dependency, permission, test-coverage) enforced automatically?
Is evidence (checks run, findings, fixes) attached to every change and retained?
Is AI coding ROI measured, with a baseline?
Related
R101 Uber SLATE (agent-produced evidence attached to every PR before review).
R103 EvoTrias (a judge the optimising agent cannot read).
R098 Stanford CS329A (verification sets quality once generation is cheap).
R102 AI-assisted performance tuning overfits (a single metric as the only verifier).