Saurabh Yergattikar

Agents Out of Bounds: Keeping Architecture Coherent - and Token Budgets Lean - Across Thousands of AI-Generated Pull Requests

Abstract:

LLM coding agents now write a large share of the code that teams merge, and they are active in roughly a 5th of public GitHub projects. The output is fast and usually correct in the small - but an agent optimizes for the task in front of it, not the global shape of the system. The cost is now measurable: adoption causally raises code complexity by about 42% and static-analysis warnings by about 30%, and the most common reason an agent's patch fails is a shallow grasp of the surrounding architecture. This talk names that widening gap between intended and realized architecture - architectural drift - and presents a taxonomy of how LLM agents produce it. 
The talk then introduces GyroCompass, an open-source guardrail framework that treats a team's architecture as a living, machine-readable artifact. It extracts that model from source across fifteen-plus languages, serves it to the agent as planning-time context (prevention), and enforces it through a deterministic commit gate that cannot be bypassed in CI (detection). On a labeled benchmark with realistic decoys, the system separates real violations from look-alikes at 100% recall and zero false positives, where a naive pattern baseline false-positives on nearly half the decoys. 
A 2nd thread runs through the talk: architecture-as-context is also a token-efficiency story. When an agent is handed a clean architectural model upfront, it stops burning tokens re-deriving structure and stops generating drift that triggers expensive multi-turn rework loops. This places GyroCompass alongside the broader context-engineering movement - operating at the architectural layer. The session closes with what reliable guardrails actually demand of a team, why the deterministic gate (not the prompt) is the binding guarantee, and what remains an open research frontier.

Profile:

Saurabh Yergattikar is a Technology Lead and Member of Technical Staff-2 at eBay Inc., Out of Total 14+ Years of career , recent 8+ years he has built core financial infrastructure serving roughly multi-million count of sellers - including eBay's first real-time fee-netting system, the Billing 2.0 platform migration, and the Seller Financing Platform across the US, UK, and Germany. His 14+ year engineering arc spans industrial IoT and cybersecurity at GE Digital, large-scale financial systems at eBay, and independent research into the security and governance of AI agent systems. 
He is an active contributor to open-source AI safety infrastructure, including SAFE-MCP (a Linux Foundation / OpenSSF project) and his own tools GyroCompass and ShieldMCP. His research on guardrails for LLM coding agents and runtime MCP security has been accepted at venues including ACL and IEEE conferences. He serves as a reviewer and program-committee member across ACM CCS, AIES, IEEE, and EMNLP, and judges hackathons including Stanford TreeHacks and the MIT Global AI Hackathon.