Agent Fix: New Agentic Workflow & Claude Model Upgrade
Snyk upgraded Agent Fix to use Claude models with enhanced agentic workflows, including agentic retries that detect and correct deviations from security best practices, dynamic few-shot prompting for secure examples, and full language coverage across all Snyk Code languages. Performance improved significantly with Sonnet and Opus models showing 10+ percentage point gains on Snyk's Golden Test benchmark.
New Model & New Architecture
We're happy to announce we're upgrading Agent Fix to use the Claude family of models enhanced by Snyk's tooling and intelligence. This move delivers the following major improvements:
Security & Functional Enhancements
Agentic Retries: Our new workflow now detects where code suggestions deviate from security best practices. Instead of discarding the result, the system analyzes the failure and injects tailored guidance into the agent's subsequent attempts.
Dynamic Few-Shot Prompting: We now use the same training set used to fine-tune our internal model to dynamically provide secure fix examples for the new model to follow.
Expanded Support
Full Language Coverage: We will enable support for all Snyk Code languages on Day 1, removing previous limitations on language availability.
Comprehensive Rule Support: AI-powered fixes are now available for all supported rules and vulnerability types across the platform.
Measurable Impact
Golden Test Benchmark: Both Sonnet 4.6 and Opus 4.6 saw improved performance against Snyk’s Golden Test benchmark (72.4% to 82.5% and 74.6% to 85.4% respectively) with this new architecture vs. the models on their own.
Check out the blog for more details. This update started rolling out on May 26th and will reach 100% by end of day on May 28th.
Source: original entry ↗