Bug Fixing is an ETL Problem

The hardest part of fixing bugs is not writing the code. It's identifying the source of the problem through an iterative search over your data.

The code is the source of all bugs, but even in today's bionic, LLM-wielding world, merely reading the code is rarely sufficient to identify all bugs. Instead, we rely on data derived from code execution to guide us to the source of unexpected behavior.

And boy, do we have a ton of data derived from code execution: Sentry, Datadog, Stripe, Supabase, and countless other SaaS products are all harboring some exhaust from recent executions. And there's always more derivative data waiting to be produced whether by adding instrumentation, attempting to reproduce behavior, or by bisecting your code.

The derived data is like a stained glass window, all refracting the original data in unique ways, and combining together to tell a story. Piecing that data together and interpreting the story, the ETL process, that is the hardest part of fixing a bug.

And this is a challenge uniquely suited for LLMs! Many successful LLM use cases share this pattern:

• Perplexity / Glean: searching over data and answering a query
• Coding agents: searching over existing code to answer a query
• Customer Support agents: searching over documentation & customer data to answer queries

What if bug investigations looked like this too: LLMs orchestrating the search over code, logs, traces, and database state, then delivering a concrete fix path? Browser and coding sub-agents supplementing the search with data generated from reproductions, bisects, and added instrumentation?

Think about your last P0 incident. How many tools did you jump between? How long did it take to correlate a user's bug report with a Sentry error, find the relevant log lines, trace the request through your services, and identify the code path that caused it?

We think that process should take minutes, not hours.

At qckfx, we're building AI agents for bug investigation that treat bug investigations as data pipelines, not a coding problem. We have launched with integrations with Github, Slack, and Sentry and are quickly adding more. If you're ready for AI-native bug fixing, you can try it free at https://qckfx.com or contact [email protected] with any questions or just to share your thoughts.