Guide, Verify, Solve — Anirban Chatterjee, Sonar
A Sonar talk argues that coding-agent gains need repeatable verification, citing a three-month productivity spike alongside persistent static-analysis warnings and complexity.
A GitHub-project study presented by Sonar found an AI-associated productivity increase that faded after about **three months**, while static-analysis warnings and complexity persisted. A separate human study cited in the talk found participants followed correct AI advice **92.7%** of the time and incorrect advice nearly **80%** of the time.
Treat verification as part of the agent harness, not a final human glance. The proposed **Guide–Verify–Solve** loop encodes constraints first, runs repeatable computational and LLM-based checks, then gives the agent tools to remediate findings before proceeding.
A GitHub-project study presented by Sonar found an AI-associated productivity increase that faded after about **three months**, while static-analysis warnings and complexity persisted. A separate human study cited in the talk found participants followed correct AI advice **92.7%** of the time and incorrect advice nearly **80%** of the time. Treat verification as part of the agent harness, not a final human glance. The proposed **Guide–Verify–Solve** loop encodes constraints first, runs repeatable computational and LLM-based checks, then gives the agent tools to remediate findings before proceeding. This was a vendor presentation, and the material does not establish that AI caused the observed project trends. It also provides no comparative results showing that Sonar's specific products outperform other layered test, analysis, and review setups.
This adds a concrete control loop to warnings that agent throughput can outpace software quality: constraints and verification must shape each iteration, not appear only at final review. The cited persistence of warnings narrows productivity claims to short-term output rather than maintainable progress, while the high compliance with incorrect advice shows why an unaided human glance is an unreliable safety boundary.