Conference Presentation, Keynote, Other
Anish Agarwal, Traversal: Production Software Keeps Breaking, and It Will Only Get Worse
- Software generated by AI systems is projected to comprise approximately 80% of total code by next year, driving increased system complexity that reduces engineer context and escalates troubleshooting difficulties.
- Future AI agent operations will predominantly focus on troubleshooting rather than system design unless automation of troubleshooting workflows is implemented, as traditional AIOps models fail due to high dynamics and false positives in large, changing systems.
- Current limitations include petabyte-scale data exceeding LLM memory capabilities, deprecated runbooks for complex incidents, and enterprise war rooms requiring days and dozens of personnel to aggregate sufficient context for root cause analysis.
- A proposed solution utilizing swarms of AI agents combined with causal machine learning, reasoning models, RAG, vector search, and code agents aims to rapidly identify specific code changes, reduce war room sizes from 50 to 10–15 people, and improve efficiency beyond capabilities existing six months ago.
- Strategic goals involve expanding from root cause analysis and remediation to encompass alerting, preventative measures, and post-mortems, with the underlying objective of enlarging data volume while shrinking the scope of the root cause.
- The technology addresses cross-industry needs in security, networking, product analytics, and business operations, supported by active hiring in New York City for talent from firms like Citadel Securities, AI product engineering, and top universities.