The poc in this repository serve as a baseline proof-of-concept. They demonstrate that using runtime evidence to build AI context is logically sound and highly effective.
However, transitioning from a conceptual demo to a production-ready AI Agent requires solving a series of hardcore engineering challenges. Real-world codebases are chaotic, polyglot, and fragile.
To validate that this paradigm truly scales to the enterprise level, I have independently developed a fully functional, production-grade implementation of this entire architecture. While that complete system currently remains my private codebase, the engineering blueprint below outlines the core capabilities I have built to make it work:
-
Polyglot / Cross-Language Support
Modern systems are rarely single-language. A true runtime-context agent must seamlessly instrument and gather unified evidence across diverse tech stacks simultaneously. -
Advanced Scenario Library Management
Moving beyond hardcoded prompts requires an AI-driven scenario library. This involves AI-based scenario search and matching, mapping specific templates to functional workflows, and orchestrating the exact execution plans required for each template. -
AI Agent "Skills" Integration
To execute complex workflows autonomously, the AI Agent must be equipped with specific, actionable Skills (tool-calling capabilities). The agent relies on these Skills to automatically analyze target projects, resolve dependency trees, and safely inject instrumentation libraries without breaking the build. -
Precision Instrumentation Control
- Blacklist/Whitelist Management: Fine-grained control over which packages or methods are instrumented to minimize performance overhead and context noise.
- Incremental Instrumentation: Dynamically instrumenting only what has changed or what is strictly necessary for the current AI task.
- Instrumentation Recovery: Safe, automated rollback to restore the target project to its pristine state after evidence is collected.
-
Deep Contextual Tracing (e.g., SQL Binding)
Code execution is only half the story. The system captures external interactions—such as triggered SQL statements—and weaves them directly into the runtime context for the AI model to reason about. -
Centralized Web Console
A dedicated Web UI is essential for managing the AI Agent, monitoring active instrumentation tasks, configuring scenario templates, and visualizing the gathered runtime evidence.