Frontier research for a superintelligent SRE
Production systems push the limits of today’s AI. Building agents that can understand and operate these vast, interconnected, constantly changing environments requires advances in harness optimization, continual learning, and reinforcement learning.
Why production is one of AI’s hardest frontiers
AI agents have made significant progress in static, single-repository environments. Enterprise production demands a different level of reasoning, learning, and evaluation.
See our research applied →
Scale & Complexity
Enterprise production spans vast, interconnected systems generating petabytes of telemetry, demanding AI that can identify and reason over what matters.
Causation vs. Correlation
Identifying root causes requires distinguishing causes from symptoms across complex system dependencies.
Constant Change
Production environments evolve continuously, requiring agents to maintain up-to-date system knowledge and reason without hindsight.
Latency
AI agents must reason across massive amounts of data and reach accurate conclusions in seconds or minutes, not hours.
Heterogeneity & Ambiguous Ground Truth
Every enterprise operates differently, and incidents rarely have definitive answers, making generalization and evaluation difficult.
High-Stakes Decisions
Production failures can have enormous consequences, demanding both speed and accuracy from AI agents.
How we are pushing the frontier
Data Processing & Production Modeling
Building agent-native data indices and RL training environments to model complex production systems and enable superhuman SRE performance.
Autonomous Agent Architecture & Optimization
Developing and optimizing agent architectures that reason about cause and effect, navigate complex systems, and make reliable decisions through harness optimization and ensembling.
Evaluation & Benchmarking
Developing rigorous evaluations with reliable ground truth and representative production tasks to measure and improve agent accuracy, reliability, and efficiency.
Continual Learning
Improving agent performance over time through reinforcement learning and distillation to adapt to evolving production environments.
Papers
Benchmarking AI for real-world production
ORCA-bench
What SWE-bench did for software engineering, ORCA-bench does for on-call: rigorously evaluate frontier models on production root-cause analysis. Developed with researchers at Columbia University and Cornell Tech.
Traversal-bench
An internal benchmark comparing Traversal against frontier models with direct tool access on real, high-severity enterprise production incidents.
Academic rigor. Production-scale challenges. Startup speed.
Deep Academic Roots
Research has been foundational to Traversal from the start. Three of our cofounders hold PhDs in machine learning, and two are professors. Today, that academic rigor extends across a broader team of accomplished PhD researchers and research engineers.
Research Meets World-Class Engineering
AI researchers work closely with distributed systems engineers, infrastructure specialists, and SREs. This tight feedback loop enables rapid experimentation, rigorous evaluation, and faster iteration in real production environments.
