Traversal 完成 1.2 亿美元 B 轮融资，估值 10 亿美元，扩展 AI SRE

Traversal announced a $120 million Series B round at a $1 billion valuation on November 3, 2026, to accelerate its work toward self-driving production systems and expand hiring across research, engineering and go-to-market roles.

Kleiner Perkins, Redpoint Ventures and Khosla Ventures co-led the round, with participation from Sequoia Capital, Battery Ventures and Amex Ventures. Earlier backers Kleiner Perkins and Sequoia increased their investments. Total funding now stands at $173 million, following the $48 million Traversal announced when it emerged from stealth in June 2025 and a strategic Amex Ventures investment in December 2025.

Traversal began with incident root-cause analysis, drawing on three co-founders’ academic research in causal machine learning and reinforcement learning. The central premise is that an incident generates many simultaneous signals, so a plausible correlation cannot establish why a system failed. The Production World Model converts raw telemetry, code, deployment history and organizational knowledge into a continuously updated, AI-readable representation of a production environment. The Causal Search Engine is designed to test thousands of hypotheses in parallel, follow dependencies and timing across systems, and converge on causes supported by evidence.

Traversal formally introduced those two systems in March 2026 after nearly two years of research and engineering. Workers now act as proactive, long-running agents that gather context, investigate and assist remediation when incidents or alerts occur. Traversal MCP makes that production intelligence available to coding agents, connecting operational evidence with development work aimed at preventing incidents.

The longer-term goal is software that detects problems, understands their causes, takes corrective action, verifies that the action worked and prevents recurrence. That roadmap makes the round an expansion of an existing investigation product as well as a bet on more autonomous operations.

Traversal reports that monthly investigations grew more than 100-fold during 2026, from 513 to nearly 73,000. Roughly two-thirds of product usage now extends beyond root-cause analysis into alert triage, release risk, CMDB validation, security, customer-journey health and production-support questions. The mix points to broader use in ongoing operational decisions alongside post-incident diagnosis.

In an accompanying funding release, Traversal reports ninefold year-to-date growth in annual recurring revenue, more than 150,000 customer investigations in the most recent quarter and support for over 4,000 engineers. Nubank, Harvey, DigitalOcean and PepsiCo are among the enterprise environments using the platform. Nubank CEO David Vélez described his company’s deployment as only five weeks old and expanding in waves as results hold up, an early-stage qualification to the adoption figures.

Benchmark Evidence and Its Limits

The technical difficulty is documented separately in ORCA-bench, a paper by researchers affiliated with Cornell Tech, Traversal and Columbia University. It contains 1,079 root-cause-analysis tasks built around six days and more than 50 GB of metrics, logs and traces from an OpenTelemetry-instrumented system with 19 microservices. Agents used real observability interfaces and source code while tasks varied report specificity, detection delays and co-occurring faults.

Across five evaluated general-purpose agents, the paper reports a best root-cause accuracy of 25.3% on medium-difficulty tasks and 10.0% on hard tasks. Hallucinated implausible root causes appeared in 7% to 40% of incident reports, depending on the model. Removing source-code access reduced accuracy and increased hallucinations for every model, reinforcing the finding that fluent code ability does not translate directly into reliable on-call diagnosis.

Those scores do not measure Traversal’s own production performance. The authors caution that the evidence comes from a controlled public testbed spanning six days, with each task handled separately. Whether the results carry over to bigger, changing proprietary environments remains untested. The benchmark therefore defines the research gap facing autonomous on-call work; it does not validate the company’s self-driving-production roadmap.

Traversal’s reported adoption and the benchmark results frame different stages of the same ambition. The current business is built around investigations and broader production questions; the open test for self-driving production is whether its causal system can take corrective action, verify the outcome and prevent recurrence across complex customer environments.