Forensic AI Agents: Accelerating Sprints Through Non-Coding Codebase Audits

Two Agents, One Codebase, Zero Collisions

The ORCHESTRATE marketing platform utilized a dual-agent system where one agent performed TDD while a second forensic agent audited the codebase. This forensic agent successfully identified a critical database mismatch involving an empty orchestrate.db file before the building agent reached the relevant tickets.

Why This Matters

In complex AI-driven systems, the gap between services compile and platform operates often hides silent failures in initialization chains or environment configurations. Forensic auditing prevents the building agent from wasting cycles on 30-60 minute diagnostic tasks per ticket, transforming debugging from a reactive bottleneck into a proactive research trail that maintains sprint velocity.

Key Insights

• Database Split-Brain: A mismatch between orchestrate.sqlite (500KB with 38 tables) and orchestrate.db (4KB and empty) caused failures across 15 endpoints.
• Silent Dependency Failures: The platform’s V3 API endpoints returned HTML instead of JSON because better-sqlite3 initialization was swallowed by try/catch blocks in lazy dynamic imports.
• MCP Tool Validation: The forensic agent identified that 102 registered MCP tools on port 3848 provided a more production-realistic validation path than the broken REST API for knowledge graph services.
• System Integrity: Audit of production data confirmed a stable LinkedIn pipeline with 554 posts and a healthy 60-second scheduler tick interval.

Working Examples

Example of lazy dynamic imports with try/catch blocks that silently swallow failures when database files are missing.

try {
const { MoeRegistry } = await import('../dist/services/moe-registry.js');
const db = new Database(path.join(DATA_DIR, 'orchestration.db'));
// ... register 28 MOE endpoints
} catch (err) {
console.warn('⚠ MOE routes not available:', err.message);
}

Practical Applications

• Use Case: A forensic agent scans ahead of feature development to identify missing database files like sources.db or provenance.db. Pitfall: Silent error swallowing in initialization scripts leads to unmounted routes that are difficult to trace during active feature building.
• Use Case: Validating complex AI services like knowledge graphs through MCP tools instead of REST layers. Pitfall: Over-reliance on REST API testing for internal AI logic can mask operational readiness of the underlying tool-calling infrastructure.

References:

• https://dev.to/tmdlrg/the-agent-that-doesnt-write-code-how-a-forensic-ai-assistant-changed-our-sprint-18d