Spiros Xanthos has spent more than two decades building systems to help engineers monitor and troubleshoot complex software. Yet the most punishing part of the job never changes: When something breaks, on-call engineers are on the hook to fix it, even if it’s the middle of the night.

While overseeing cybersecurity and data platform Splunk’s software monitoring teams, Xanthos saw this strain firsthand as his teams integrated newly acquired tools while trying to keep its existing systems running.

“[It] was so hard on our site reliability engineers,” Xanthos says. “Over a period of six months, we lost 90% of them. It was a complete burnout.”

That experience led Xanthos and his cofounder Mayank Agarwal to launch Resolve AI in 2024 to automate the painful process of responding to production issues. Traditionally, teams rotate who has to be on call, where developers are paged when systems fail. With Resolve’s system, multiple agents investigate when a problem is flagged. They orchestrate a host of AI models and tools to analyze logs, metrics and test different hypotheses like traffic spikes or faulty code. If Resolve can automatically fix the issue, it will. If it’s more complex, the system will present a recommended solution to a human engineer, who can review and approve it.

“We're still in this human-in-the-loop process,” Xanthos says. “But we're moving more to this concept of human on the loop – a human can review what Resolve does, but doesn’t have to pause and decide.”

Resolve’s customers, like Coinbase, DoorDash and Salesforce, have seen dramatic improvement in the time it takes to fix an incident and investigate why it occurred. According to Resolve, DoorDash reduced investigation time from roughly 40 minutes to about one minute, while the time to identify the root cause dropped by up to 87%. Coinbase resolved incidents roughly 72% faster when engineers used Resolve compared to working alone, Xanthos says.

That convinced investors to pour an additional $40 million into Resolve, a follow-on to its $125 million Series A round it raised just months ago, led by Lightspeed Venture Partners at a $1 billion valuation. This extension round brings Resolve’s total funding to over $190 million and valuation to $1.5 billion. The startup was featured on Forbes' inaugural AI 50 Brink list.

According to Rahul Mehta, an investor at DST Global who led the deal, one customer found that Resolve was able to identify issues that other tools failed to detect, in addition to reducing investigation time.

“Production systems are fairly complex and they are mission critical…which is why you need high accuracy and reliability,” says Mehta, who made early investments in Facebook and Spotify and has been a mainstay of the Midas list since 2019. “Given the low margin of error, it needs to be a perfect system pretty much overall, and AI has a real role to play.”

Resolve now has around 140 employees, including more than 20 people it hired away from Google DeepMind. While pricing isn’t publicly disclosed, the company sells its product through a credit-based model tied to the amount of work the platform performs rather than a fixed per-use fee.

Xanthos and Agarwal, both 43, first met as PhD students at the University of Illinois Urbana-Champaign in 2004. Though Xanthos eventually left the program without finishing, the two remained connected and later helped create OpenTelemetry, an open source project that standardizes how companies collect data about how their software is functioning. Building on that concept, in 2018 the pair cofounded Omnition, a startup that helped companies track how user actions move through different parts of their system, which Splunk acquired in 2019 for $52.5 million, according to PitchBook.

Resolve is part of what Mehta describes as a nascent category of “AI for production” tools, designed to automate how companies diagnose and fix issues in incomplete and complex software systems. It is also part of a broader wave of startups trying to automate different aspects of software development and operations, from coding assistants like Cursor, valued at $29.3 billion, to the $580 million AI cybersecurity firm Depthfirst, which builds its own models to fix security vulnerabilities. Similarly, AI cyber startup Corridor recently raised funding at a $200 million valuation to catch code errors before an attacker does. Then there’s the 800-pound gorilla in the room: Anthropic’s Claude Code.

But handing over core aspects of software production to AI introduces a new set of risks. As engineers rely more and more on tools like Claude Code and Codex to write the bulk of their code, they may only partially understand the codebase they’re working in, raising concerns around security and reliability.

The consequences of limited oversight are already starting to emerge. For example, in March, Amazon faced disruptions after its coding assistant Q helped generate a code change that was pushed into production, causing incorrect delivery times for customers, 120,000 lost orders and 1.6 million website errors, according to Business Insider. The Financial Times reported in February that a 13-hour disruption to Amazon Web Services in December was linked to changes made by its Kiro AI tool. Amazon said the incident was limited in scope and it has since implemented additional safeguards.

In addition, the open source project LiteLLM experienced a major cybersecurity attack in March. Some observers suggested that messy vibe coding was to blame for the breach, including Andrej Karpathy, a cofounder at OpenAI who now runs education startup Eureka Labs.

Resolve’s founders say security is top of mind. Agarwal, the company’s CTO, says he personally speaks to every customer’s security team and provides them with controls to see what actions are taking place in their codebase.

But the bigger challenge is just keeping up with the pace of change right now, he says. “Customers have come to expect a magical experience and living up to that requires constantly rearchitecting and reevaluating the right approach.”

That’s why Resolve is now building out its own internal AI lab focused on building AI tools so that LLMs can more effectively tackle the complexity of production environments, where incidents can span multiple teams and tools. The company recently recruited Dhruv Mahajan from Meta’s Superintelligence group to run the lab and serve as Resolve’s chief AI scientist. Mahajan, who holds a PhD in computer science from Columbia University, will lead efforts to build AI systems that can operate across entire production workflows end-to-end.

While Resolve acknowledges the technology isn’t yet accurate enough to replace humans fully, Xanthos expects Resolve to be able to address most production issues automatically by the end of next year.

“There’s no developer out there that won’t be happy to give this task to AI,” Xanthos says.

© Forbes