120 million dollars to fix the code that AI helps produce

Valued at $1 billion, Traversal develops agents capable of diagnosing software failures and finding their root causes.

Writing code is getting faster. Understanding why it stops working in production does not necessarily keep pace. As companies delegate more development to AI assistants, their systems accumulate changes, dependencies, and behaviors that teams must then monitor. Traversal aims to automate part of that work, with a $120 million funding round that brings its valuation to $1 billion.

This Series B, co-led by Kleiner Perkins, Redpoint Ventures, and Khosla Ventures, also includes participation from Sequoia Capital and Battery Ventures. Founded in 2023 by Anish Agarwal, Raaz Dwivedi, Raj Agrawal, and Ahmed Lone, the U.S. startup is developing a system of agents specialized in Site Reliability Engineering (SRE), the discipline responsible for keeping IT services available and running properly.

According to information gathered by Forbes, Traversal’s agents took part in more than 150,000 investigations last quarter. The company says it multiplied its annual recurring revenue ninefold over a six-month period, without disclosing the absolute amount. Among the organizations using it are Nubank, Harvey, DigitalOcean, and PepsiCo, as well as several major U.S. companies. An outage can have thousands of apparent causes

When a service becomes unavailable, technical teams generally have a substantial amount of signals at their disposal: execution logs, alerts, performance metrics, deployment histories, and traces of communication between applications. The challenge is determining which ones actually explain the incident.

A slow database can trigger a cascade of errors across multiple applications. Monitoring dashboards then simultaneously flag dozens of symptoms, without necessarily revealing the initial failure. A language model can offer a plausible explanation based on this information, but temporal correlation alone is not enough to establish a causal relationship.

That distinction is what the Causal Search Engine is built on. Developed from the founders’ research in causal learning and reinforcement learning, it tests different hypotheses against the real dependencies of systems, the sequence of events, and the changes introduced before the outage.

Investigations can be carried out simultaneously across a large number of leads. A hypothesis is ruled out if it does not match the relationships between components or the observed spread of errors. The diagnosis must therefore rest on a chain of verifiable evidence rather than on the mere linking of unusual signals.

However, this approach requires a representation of the infrastructure that is complete enough to distinguish the consequences of an outage from its origin. A living map of infrastructure

Traversal’s second component, called the Production World Model, brings together information scattered across observability tools, code repositories, deployment histories, and incident reports.

Instead of consulting each source separately, the agents work from an up-to-date representation of services, their dependencies, and their behavior. A change in one application can be linked to the components it calls, the operations performed on databases, and prior incidents with similar characteristics.

This representation is enriched by the operational knowledge accumulated by teams, including information that is rarely recorded in official documentation. The system thus seeks to reconstruct the actual relationships between the elements of an infrastructure, which may differ from the theoretical diagrams established when it was designed.

According to Traversal, it took nearly two years of research to develop this architecture, whose two main components were presented in March 2026.

The claimed usage volumes point to rapid adoption: the number of monthly investigations reportedly rose from 513 to nearly 73,000 over the course of the year. These figures measure platform activity, not the number of outages automatically resolved. The company’s official communication does not provide an independent protocol for evaluating diagnostic accuracy across all of its customers. Agents that step in before engineers do

Traversal’s activity is no longer limited to finding the root cause of service interruptions. About two-thirds of uses now involve other operations, such as alert triage, risk assessment before a deployment, IT inventory verification, or analysis of how customer journeys are performing.

Since September, Traversal Workers have been able to launch investigations when an anomaly is detected, without waiting for an engineer to submit a request. These persistent agents gather the necessary information, examine possible causes, and prepare remediation actions.

The system also extends to development tools through Traversal MCP. Coding assistants can consult information from the production environment to better understand the potential consequences of a change before it goes live. Observations gathered during past incidents thus become usable while new features are being built.

The ambition is to connect two stages of software development that are traditionally separate: producing code and ensuring it works after deployment. Companies could then have a loop in which problems encountered in operations directly help prevent similar mistakes in subsequent development.

This vision, described by Traversal as self-driving production, nonetheless goes beyond currently documented capabilities. Detecting an anomaly, proposing a diagnosis, and preparing a fix do not guarantee that an agent can operate without supervision on a critical infrastructure. Permission to make changes, error handling, and verification of fixes remain decisive questions. Growth funded before full autonomy

The announced funding round is primarily intended to support the research, engineering, and business development teams. Anish Agarwal plans to double headcount over the next six months in order to expand the agents’ capabilities and their deployment in large enterprises.

The acceleration comes in a market where