At a glance
WHAT IT’S REALLY ABOUT
AI-written code boosts outages; Traversal builds AI reliability engineer
- Agarwal argues the biggest hidden risk of AI coding is loss of system context, making outages harder and more expensive to debug as code volume explodes.
- Traversal positions itself as an “AI site reliability engineer” that investigates incidents by querying observability data and producing evidence-linked root-cause hypotheses.
- The company learned that an MVP that works in small environments can fail catastrophically in enterprise settings, prompting a redesign from “clever prompts” to computation-heavy inference workflows.
- He frames modern reasoning models as well-suited to “detective story” tasks—connecting many symptoms to a single culprit—mirroring incident response.
- The conversation also covers founder lessons: iterate at the edge of model capability, build resilience through failures, and choose investors/mentors/teammates carefully.
IDEAS WORTH REMEMBERING
5 ideasAI coding creates a context deficit that amplifies outage risk.
As tools like Copilot/Cursor generate more code, fewer humans hold end-to-end understanding of systems, so debugging becomes slower and more chaotic when incidents occur.
Incident response is a high-ROI target because downtime is massively costly.
Agarwal cites enterprise downtime costs on the order of hundreds of billions annually, and describes how incidents often balloon into large multi-team war rooms without fast attribution.
An effective AI SRE must provide traceable evidence, not just answers.
Traversal’s analogy is Perplexity-style citations, except the “sources” are logs, metrics, traces, code, and config links inside the customer’s observability stack.
Enterprise-scale complexity can collapse MVP performance to zero.
Traversal saw ~90% accuracy with small companies but dropped to 0% with a large provider (DigitalOcean), illustrating that production robustness and data scale are fundamentally different from demos.
Design for computation and inference, not founder “creativity in prompts.”
They recovered performance by re-architecting to exploit what models are good at—systematic inference over large evidence sets—rather than brittle handcrafted reasoning.
WORDS WORTH SAVING
5 quotesAI is gonna write so much more code. No one really understands all of it. No one has full context. No team has full context about what's happening because it's such a complex system. When software breaks, it's gonna be really difficult to troubleshoot it, right?
— Anish Agarwal
The thing that changed-- And that was the time when everything with Chat GPT was, was happening, and so it just felt like something incredible has happened in the world, and it's like a once in a lifetime thing where the world has fundamentally changed. People don't even realize it. It just felt like this almost religious experience as to what was happening in the world.
— Anish Agarwal
Creating an MVP is easy, but creating a production system that works in complex environments is really hard. And so I think once you're not confused an MVP with a production AI system, those are like two very, very, very different things.
— Anish Agarwal
Our accuracy went to 0%, which is very difficult to see. It was a tough week.
— Anish Agarwal
Over time, all engineers will be doing will be troubleshooting, which will be sad in my opinion. Like they should be doing the most-- We as a, as engineers should be doing the most creative work, right?
— Anish Agarwal
High quality AI-generated summary created from speaker-labeled transcript.
