How One of AI’s Fastest-Growing Companies Pairs Its Own AI SRE with Traversal Workers for Faster Remediation

At a Glance

Harvey is one of the fastest-growing companies in AI, building the platform that law firms and enterprises now use to do legal work. Thousands of pull requests land daily, and product enhancements reach customers in days. Where Traversal's enterprise customers bring distributed scale and regulatory complexity, Harvey brings the AI-native test: a company building at the forefront of AI, with a production environment that changes by the hour.

Harvey’s engineers built Spectre, a highly capable internal background agent they use for everything from shipping pull requests to investigating alerts and SEVs. Spectre already set a high bar. Driven by its commitment to enterprise customers, Harvey continually looks for opportunities to raise the bar on reliability. So they evaluated Traversal live on real production SEVs, with Harvey’s own engineers doing all the scoring.

Traversal delivered a bullseye or directionally correct root cause on 92% of SEVs, with no misleading answers and a median time to final answer of 10 minutes.

Today, Spectre and Traversal both join every Harvey SEV, where they complement and cross-validate each other. Harvey added Traversal for four reasons:

• Accurate root cause analysis on live production SEVs.
• Workers that engage proactively and stay with a SEV until it closes.
• High signal and low noise in SEV channels.
• A team that executes and iterates quickly in a field that changes month to month.

"I’m impressed by Traversal’s RCA accuracy, speed, and signal-to-noise ratio, and how quickly their team acts on feedback—that matters a lot for a fast moving field like AI SRE. Spectre already set a high bar, and Traversal really earned its place to become a critical part of our reliability strategy."
Zhiyu Chen, Engineering Leader, Harvey

The Challenge

Harvey is the AI platform leading law firms rely on every day. Reliability is critical to Harvey. In order to minimize interruption or degradation to the services that they provide, Harvey declares SEVs for any system anomaly before they have any external impact. Shipping at Harvey’s pace makes that harder in three ways.

• Velocity Outpaces Context. With thousands of PRs landing daily, no single engineer can hold the full picture of what changed, where, and why. Harvey wanted any engineer on the team to be able to own a SEV.
• Noise Kills Adoption. An AI SRE that floods the channel with hedged, speculative updates gets muted, however accurate it is. Harvey made signal-to-noise a formal evaluation criterion.
• Reactive Bots Can’t Keep Pace. Spectre already investigated SEVs automatically and followed new evidence; Harvey required the vendor solution to meet that workflow: one that decides for itself when to engage, picks up SEVs on its own, and stays with them until they’re closed.

Our Evaluation

Traversal connected to Harvey's observability systems via agentless, read-only access. Most importantly, Harvey's engineers never wrote a single markdown file. Other AI SRE vendors depend on hand-written context via markdown files about services, dependencies, and runbooks, and at thousands of PRs a day, that context goes stale fast; Traversal learned Harvey's environment on its own and began investigating right away.

Most AI SRE evaluations backtest on historical SEVs or run in shadow mode, which typically has a large gap to live production SEVs. As a result, Harvey tested live, on real production SEVs, in its actual SEV channels.

Harvey’s engineers also scored the evaluation themselves. They graded responses across 24 SEVs on a 0–10 scale, from bullseye to misleading, and measured time to first and final answer. The criteria covered accuracy, latency, and user experience, including the signal-to-noise ratio of each response.

Traversal’s Impact at Harvey

Root cause accuracy:

• A bullseye or directionally correct root cause on 22 of 24 SEVs (92%).
• No misleading answers. Traversal never sent the team in the wrong direction.

Latency. Traversal's median time to a final answer was 10 minutes.

What Changed for Harvey's Engineers

Traversal Workers are proactive, always-on AI SRE agents that pick up SEVs and investigate them end to end autonomously, like an engineer on the team. For Harvey, that was non-negotiable; a Slack bot that waits to be asked still leaves engineers doing the hard part themselves.

Before: Spectre was helping Harvey’s engineers investigate production SEVs and gather context. Difficult SEVs could still require substantial service-specific knowledge and additional investigation, leading to mitigation delays.

After: With Traversal alongside Spectre: human engineers have another source of evidence. On some SEVs, Traversal identifies the cause more accurately or reaches a useful conclusion sooner; on others, both systems complement and cross-validate each other. Engineers can compare their conclusions, investigate disagreements, and validate proposed mitigations against the underlying evidence. Together, the agents make the investigation context more accessible. Even newly joined engineers can get up to speed from the channel and move straight to mitigation.

Why Production RCA Needs Specialized Architecture

General-purpose agents are powerful for shipping code, but root cause analysis is a different problem. It depends on knowing how everything in production connects, and on keeping that knowledge current.

That's what Traversal's Production World Model™ does. It maps how Harvey's services, dependencies, deploys, and infrastructure connect into a single AI-readable model, built automatically from telemetry, code, documentation, and tribal knowledge. As Harvey ships thousands of changes a day, the model updates itself, with no markdown files and no manual mapping.

Faster Mitigation, More Time to Build

Reliability is a top priority for Harvey. Their evaluation showed that even with a strong internal agent, adding accurate, proactive root cause analysis helps the team mitigate faster and limit customer impact. And with Traversal handling the specialized work of RCA and preventing further SEVs, Harvey’s engineers get more of their time back for building the product law firms rely on. For any team shipping at AI speed, that’s what reliability has to look like: faster answers when systems are affected, and more time to build when they’re not.

To see how Traversal can work in your environment, book a demo today.