Descriptions:
Tariq Shaukat, CEO of Sonar, argues at the AI Engineer conference that code verification — not generation — is the defining bottleneck in agentic software development. Drawing on Sonar’s benchmarking of over 4,000 programming problems across state-of-the-art models, he demonstrates that while models like GPT-5 score well on functional correctness, they consistently produce complex, buggy, and insecure code. Real-world failures at KPMG and major law firms illustrate the enterprise consequences.
Shaukat presents benchmark data showing coding agents (via the Mythos leaderboard from Meter) can complete tasks equivalent to 16–18 hours of human work at 50% accuracy, but drop to roughly 3.5 hours when the accuracy bar rises to 80% — a threshold one customer CTO still flagged as a performance-review problem.
The talk introduces Sonar’s “Agent Centric Development Cycle” (AC/DC) framework, which bakes verification into the agentic loop at every stage rather than treating it as a downstream afterthought. Shaukat also announces the launch of Sonar Vortex, a new product focused on giving agents architectural context and explicit coding constraints — helping them understand both the existing codebase and where the organization wants to go.
📺 Source: AI Engineer · Published July 20, 2026
🏷️ Format: Keynote Launch
