Traversal raised $120 million at a $1 billion valuation on 3 November, taking its total funding to $173 million as it expands AI agents for diagnosing failures in enterprise software, the New York company announced in a Business Wire release. Its agents focus on site reliability engineering, which aims to keep production software running and accessible.
Key points
- Kleiner Perkins, Redpoint Ventures and Khosla Ventures co-led the $120 million round.
- Traversal reports ninefold year-to-date growth in annual recurring revenue and more than 150,000 customer investigations in the most recent quarter.
- Nubank says it is expanding its use of Traversal in stages after five weeks.
- On a benchmark developed with two universities, the best of five general-purpose agents achieved 25.3% accuracy on medium-difficulty root-cause tasks.
Kleiner Perkins increases its Traversal investment
Redpoint Ventures and Khosla Ventures joined Kleiner Perkins in leading the financing. Sequoia Capital, Battery Ventures and Amex Ventures also participated. Traversal said the money would support development of its technology and hiring in research, engineering and go-to-market roles. Its stated longer-term aim is to extend AI through the software development lifecycle by making production systems more resilient.
Traversal emerged from stealth with $48 million in funding in June 2025, followed by a strategic investment from Amex Ventures that December, Citybiz reported. Kleiner Perkins increased its investment in the new round. Sequoia invested through its growth fund to expand its stake.
The company’s Production World Model is designed to assemble information about a customer’s live software environment and its dependencies. Its Causal Search Engine is designed to investigate possible causes when that environment fails. Traversal began with root-cause analysis and has since added agents that gather context and assist engineers when alerts or incidents occur, Unite.AI, an AI industry publication, reported.
The Production World Model draws on telemetry, code and deployment history to maintain an AI-readable account of a production environment. The Causal Search Engine is designed to test thousands of possible explanations in parallel, following dependencies and the timing of changes. Traversal introduced the two systems in March 2026. It also offers Traversal MCP, which makes production information available to coding agents for development work aimed at preventing incidents.
Nubank expands use after five weeks
Traversal says its annual recurring revenue has grown ninefold year to date, while product usage has grown nearly 100-fold and monthly active users nearly twentyfold. Customers conducted more than 150,000 investigations in the most recent quarter, according to the company. Its platform served more than 4,000 engineers during over 1,000 serious incidents, 50,000 less serious incidents and 5,000 substantial updates to production systems.
Monthly investigations rose from 513 to nearly 73,000 during 2026, Unite.AI reported. The publication said roughly two-thirds of product usage now extends beyond root-cause analysis to work including alert triage, release risk and production-support questions. Those uses concern decisions made during routine operations as well as investigations after a failure.
Traversal names Nubank, Harvey, PepsiCo and DigitalOcean among its customers. David Vélez, Nubank’s founder and chief executive, said after five weeks that Traversal was becoming part of how the bank’s on-call engineers handled incidents. Nubank was expanding the rollout in stages as results held up, he said in Traversal’s release. Harvey chief technology officer Siva Gurumurthy said its engineers used Traversal as the company scaled its software development work into production.
ORCA-bench puts five agents through incident tasks
Traversal developed ORCA-bench with Cornell Tech and Columbia University to test AI agents on 1,079 root-cause analysis tasks. According to Traversal’s announcement, the best of five general-purpose agents achieved 25.3% accuracy on medium-difficulty incidents, dropping to 10% when incident reports were vague. Hallucination rates reached 40%.
Unite.AI reported that the underlying paper recorded 10.0% accuracy on hard tasks and implausible root causes in 7% to 40% of incident reports, depending on the model. The tasks drew on six days of data from a system with 19 microservices. Agents had access to observability interfaces and source code, while the tests varied the detail in incident reports, detection delays and concurrent faults.
The paper’s authors cautioned that the tests used a controlled public system and handled each task separately, according to Unite.AI. They reported that removing access to source code reduced accuracy and increased hallucinations for every model tested.
.jpg){kind=link}