Kolton Andrus and the art of an uneventful day

Kolton Andrus once helped investigate a drop in orders in Japan. For an engineer responsible for keeping Amazon running, a falling number could mean broken software, a failed dependency, a bad deployment. This time, the systems were working. There had been an earthquake. People had other things to do.

He told that story years later while explaining what he learned from running major incidents. It is a useful introduction to someone whose working life has revolved around failure: the first job is to understand what is actually happening. A dashboard can tell you that something changed. Deciding what the change means still requires judgment.

Andrus has spent years trying to move that judgment into the hours before an emergency. He built failure-injection tools at Amazon and Netflix, then co-founded Gremlin in 2016. The proposition sounds mischievous: introduce a carefully chosen problem into software and watch what happens. The intended result is pleasantly uninteresting. Customers continue shopping, videos continue playing, and the engineer's phone stays quiet.

The software company he promised his aunts

At ten, Andrus told his aunts he would start a software company. He was a gamer, so he imagined making games. College supplied a commercial revision: creating games looked like considerable work, and enterprise software paid better. A childhood ambition survived its first encounter with the accounts department.

He also began keeping a mental list of people he might want to work with again. That detail explains something about Gremlin's eventual beginnings. The future founder was paying attention to colleagues long before he had jobs to offer them. Talent mattered; so did knowing how someone behaved when work became difficult.

Andrus earned bachelor's and master's degrees in computer science at the University of Utah, finishing them in 2005 and 2007. His early career included software engineering at Meridias Capital and Mindshare Technologies. Those names receive less attention than Amazon and Netflix, but they place his founder story in a longer apprenticeship. He had worked inside smaller businesses before he began studying the failure modes of enormous ones.

At Amazon, he met Matthew Fornaciari, who helped build the internal tooling and would become Gremlin's co-founder. Their connection was practical and personal: they had worked together, become friends, and seen one another through difficult stretches. When Andrus later needed a partner, he already had someone whose judgment he trusted.

The person on the bridge

Amazon made him a Call Leader, a role he also held at Netflix. The title sounds like a minor distinction until a large service stops working. Then the call leader coordinates the response, weighs proposed changes, and helps decide which risks to take while other engineers diagnose and repair the system.

Technical knowledge is necessary. So is an understanding of the business. Restarting a service might restore one function and interrupt another. Waiting for more evidence has a cost, too. Andrus has described the role as an exercise in judgment, with engineers depending on someone to make a decision when certainty is unavailable.

The experience gave him a reason to care about rehearsal. Incident response is an expensive moment to learn who owns a service or how another team communicates. In a later conversation about engineering culture, he argued that teams should meet before the outage, when they can talk calmly and build trust. Software may be distributed across machines; responsibility is distributed across people. Both arrangements can produce surprises at their boundaries.

That is part of the appeal of planned failure. A test can expose a timeout that never happens, an alert nobody receives, or an assumption two teams interpret differently. It gives the technical system and the organization a chance to reveal themselves under conditions people have chosen.

A monkey with boundaries

Netflix already had a reputation for deliberately disrupting its own infrastructure. Chaos Monkey made the idea memorable. Andrus worked on a more precise question: how could engineers introduce realistic failures while controlling which requests and customers felt them?

In October 2014, he co-authored the introduction to FIT, Netflix's Failure Injection Testing platform, with Naresh Gopalani and Ben Schmaus. Earlier approaches could spread trouble farther than developers expected. FIT let them narrow the scope, beginning with a test account or device, then gradually increasing exposure as confidence grew.

The distinction is easy to miss beneath the entertaining names. Deliberate failure needs boundaries. An experiment that surprises everyone with its damage is a poor advertisement for the next experiment. FIT's contribution was to make the intervention more selective, allowing engineers to test assumptions without immediately enrolling the whole audience.

His work also brought him into collaboration with computer scientist Peter Alvaro. Together they discussed applying failure-testing research at Netflix in their 2016 “Monkeys in Lab Coats” presentation. The project connected academic reasoning about failures with the practical constraints of a working service. Their lessons included finding shared goals and adapting theory to the untidiness of real systems.

It was a revealing partnership. Once a system has many dependencies, the list of possible combinations of failure becomes difficult to explore by hand. Better tools matter, but so does choosing better questions. The lab coat in the title was doing some work alongside the monkey.

“I'm a builder at heart.”

Kolton Andrus, in conversation with Rootly

Making an undo button into a business

Gremlin turned that experience into a commercial product. Its public launch came on December 12, 2017, after nearly two years of development. Andrus announced a $7.5 million Series A led by Index Ventures and Amplify Partners. Early customers included Twilio, Expedia, Confluent, and Remind.

The launch emphasized safeguards as much as failure: an undo button, automatic cleanup, permissions, and auditing. Those details belong near the center of the story. Asking an engineer to introduce a fault means asking them to accept responsibility for the consequences. Safety controls make the invitation more credible.

There was also a difference between building a tool inside a large company and selling one to other companies. Internal infrastructure supplies assumptions about authentication, deployment, and how colleagues work. A product has to accommodate customers who arrive with different answers. Gremlin had to package the practice along with the machinery.

In a 2022 conversation with PagerDuty co-founder Alex Solomon, Andrus recalled an earlier debate with venture investors after a conference talk. He had wanted to bootstrap the company. The discussion eventually reached a practical consideration: he had five children and lived in California. The business plan had to account for a household as well as a codebase.

A founder changes seats

By 2021, Gremlin had exceeded $10 million in annualized recurring revenue. In February 2022, the company announced that Josh Leslie would become CEO and Andrus would move to CTO, leading product and engineering. More than 100 enterprise customers were using Gremlin at the time.

For a technical founder, the change carried a familiar question: where can I do the most useful work now? Starting the company had required Andrus to learn sales, recruiting, marketing, and administration alongside product development. Moving to CTO gave him room to concentrate on what engineers and customers needed.

He returned to the CEO role in 2025. Looking back in an October 2026 interview, he described the intervening years as a chance to watch another executive work, learn from the experience, and get closer to the product. There is something fitting about a failure-testing specialist treating his own job arrangement as revisable. A title could change while the underlying problem continued to hold his attention.

What does success look like when nothing happens?

AI-generated software has given that problem a new setting. In an April 2026 conversation with Sean Falconer, Andrus explored a gap between producing code quickly and understanding how the resulting system behaves. Engineers accumulate lessons from operating software. Generating more code does not automatically reproduce that experience.

For him, the question follows naturally from the earlier work: what happens when a dependency slows down, disappears, or behaves differently? A program can look convincing on the expected path and still disappoint when another service refuses to cooperate. Speed makes the need to test those conditions more pressing.

On October 7, 2026, Andrus announced Gremlin's Foresight AI. The product assigns supervised agents to analyze services, recommend and run tests, explain outcomes, and guide repairs. Gremlin says it can rerun a failed test after a proposed fix, checking whether the change addressed the weakness. Tests and remediations require user approval.

The recurring idea is evidence. A plausible repair deserves another experiment. For Andrus, automation is a way to extend a practice he has been pursuing since Amazon, with more of the preparatory work handled by software.

He has an equally persistent organizational problem: prevention is difficult to applaud. An outage gives everyone an event to remember. A weakness found and repaired beforehand leaves a quieter record. Making that work visible helps a team keep doing it.

Andrus describes himself as a gamer and a father of five. Asked what motivates him, he points to his children and to the pleasure of building things that grow over time. The ambition he announced to his aunts has acquired employees, customers, investors, and several job titles. Its preferred ending remains modest: someone clicks, the software works, and nobody has a story to tell about it.