Transforming Debugging and Root Cause Detection with Sentry

Introduction

This conversation with David Cramer, co-founder of Sentry, explores the evolving role of large language models (LLMs) in software development, both within Sentry’s product ecosystem and in personal development workflows. It covers practical use cases, limitations, and the cultural and technical shifts necessary to integrate AI into engineering teams. Cramer brings a practitioner’s perspective, focusing on applied engineering rather than academic theory, and emphasizes that AI tools must augment rather than replace the fundamentals of software engineering.

Applying LLMs at Sentry

Sentry, a developer-focused tool, captures and processes application errors with highly structured data. This structured approach has allowed them to experiment meaningfully with LLMs. Early wins included using embeddings to improve error deduplication, replacing tedious rule-based systems. More recently, Sentry has been developing "Seir," a root cause analysis tool that interlinks rich, application-specific data with LLM reasoning. While the technology remains prone to hallucinations and non-deterministic results, Seir has demonstrated moments of unexpectedly accurate debugging similar to the skill of a senior engineer who deeply understands a system.

Sentry’s AI applications focus on real-world utility over novelty, exploring areas like automated bug fixes and test generation. These initiatives hinge on the company’s moat of interconnected, high-quality data in contrast to vendors that operate at more superficial levels of system monitoring.

Eight Weeks Without Writing Code by Hand

Cramer conducted an experiment to avoid manually writing code for eight weeks, instead working entirely through AI agents. The goal was to understand the limitations, strengths, and workflow implications of agent-based development. He rejects the idea of "vibe coding" which is blindly accepting whatever an agent outputs, insisting that proper system design, specifications, tests, and code review remain central.

The experience revealed that AI output quality varies less by task size and more by pattern familiarity. Simple, common patterns are handled well, while novel or nested agent scenarios quickly degrade into confusion. Reviewing every change is essential, making the integration experience critical. Cramer found IDE-integrated tools like Cursor superior to terminal-based workflows like Cloud Code because they support incremental review and maintain human-in-the-loop control.

Testing, Documentation, and Simplicity

Tests played a central role in Cramer’s agent-driven workflow, particularly when paired with simple, predictable system architectures. He emphasizes that LLMs excel when APIs and code are easy to reason about. This principle suggests a future where complexity is reduced to make both human and machine reasoning more effective.

While humans rarely revisit documentation after onboarding, LLMs rely heavily on it. This shift makes well-structured, machine-readable documentation more valuable, encouraging practices like exposing internal and external docs via machine-consumable APIs.

Cultural and Organizational Shifts

For teams, the first barrier to AI adoption is access, both in terms of licensing and compliance. Sentry removed internal restrictions on source code exposure to accelerate experimentation with AI tools, offering company-wide access to platforms like Cursor and Cloud Code. Overcoming skepticism, particularly from experienced engineers, requires education and demonstration that effective AI use depends on supplying relevant context and documentation.

Cramer notes an indirect but significant benefit. AI tools have re-engaged senior technical talent, drawing them back into hands-on building after years of primarily managing. This renewed enthusiasm drives innovation and better decision-making at the leadership level.

Present and Future Value

LLMs already deliver tangible productivity gains in certain workflows such as boilerplate generation, testing, and repetitive API pattern creation. More importantly, using these tools builds the organizational and individual “muscle” to exploit future advances. While full autonomy remains unlikely in the near term, keeping humans in the loop ensures quality and enables incremental improvements.

Cramer’s advice to individual developers is to embrace the technology, remain skeptical of overblown claims, and focus on hands-on experimentation. Ignoring AI’s trajectory risks obsolescence both for engineers and for companies.