Innovation & Tech

AI Writes Code Fast. Fixing It Still Takes Engineers Hours

Google's 2025 DORA report, surveying nearly 5,000 tech professionals, links AI adoption to faster software delivery but reduced stability — and startups like DataAgent want AI to fix what breaks.

By Grace Kim

5 min read

AI is making software easier to build. Keeping it running is another story
AI is making software easier to build. Keeping it running is another storyThis_is_Engineering / pixabay

What's News

  • Google's 2025 DORA report surveyed nearly 5,000 technology professionals and linked greater AI adoption to faster delivery but reduced stability.
  • PagerDuty's 2026 State of AI-First Operations report found 44% of organizations require human involvement when AI fixes customer-facing systems; 43% for cross-team incident coordination.
  • DORA researchers found developers using AI saved time writing code but spent more time reviewing and verifying it.
  • DataAgent, led by CEO Ishay Yaari, deploys AI agents inside customers' infrastructure for autonomous remediation and has not released independent performance data.

Google's 2025 DORA report, based on responses from nearly 5,000 technology professionals, found that greater AI adoption correlates with faster software delivery — and reduced stability. That single finding frames the next battle in enterprise software: not writing code, but keeping it running.

A feature that takes minutes to build with AI tools can introduce problems that take engineers hours to investigate. The researchers didn't establish that AI-generated code directly causes more outages, but their data suggests companies can produce software faster without getting any better at maintaining it.

A growing group of companies is now developing AI systems that investigate failures, identify likely causes and, in some cases, take corrective action without waiting for an engineer.

Why does more code mean more failures?

Ishay Yaari, cofounder and CEO of DataAgent, a company developing autonomous remediation software, says the operational side of development hasn't kept pace with coding.

"More code, more frequent deployments, and more infrastructure changes naturally create more potential failure points," he told Fast Company.

For site reliability engineering teams, that means managing applications and infrastructure that change constantly. When an app starts crashing after an update, the cause could be faulty code, a bad configuration, insufficient computing resources, or a problem with another service. Monitoring tools spot the failure immediately. Engineers still have to dig through logs and recent changes before deciding what to fix.

Yaari says companies often respond by adding monitoring tools and collecting more data. That shows engineers what's happening — and piles up more alerts without necessarily resolving anything faster.

The pressure doesn't come solely from poor AI-generated code. Even reliable code gets hard to manage when applications, infrastructure and dependencies keep shifting.

Google's DORA researchers, in a separate analysis of AI-assisted development, found developers saved time generating code but spent more time reviewing and verifying it. Testing, feedback and established engineering practices determine whether faster development produces better results.

Takeaway: the payoff from AI coding tools depends on what happens after developers hit submit.

Can AI get the keys to production?

Traditional monitoring and observability tools help engineers detect problems and understand causes. DataAgent wants to automate the response itself. Its platform monitors applications and cloud infrastructure, flags issues and helps investigate them.

Many failures have familiar remedies: an app crashes for lack of memory, a deployment fails on a wrong configuration, a new release introduces errors that vanish on rollback. Companies already handle some of these through predefined rules. AI could extend that by examining circumstances, identifying likely causes and selecting a response.

Yaari says DataAgent's platform is built to understand both the infrastructure and the application itself — how services interact, what the app should do, how its behavior changes after an update. The aim is catching problems no rule anticipated.

But diagnosis isn't action. A rollback may restore service, yet if the update also changed a database or introduced new dependencies, reversing it can create new problems. That's why the ability to act matters as much as the ability to diagnose.

Takeaway: recognizing the fix is the easy half; letting software touch a live system is the hard half.

How much should AI be allowed to fix?

PagerDuty's 2026 State of AI-First Operations report found that 44% of surveyed organizations required human involvement when AI fixed problems affecting customer-facing systems. Another 43% required it when AI coordinated incident responses across teams.

Yaari agrees some decisions should stay with people. Deleting large amounts of data, disabling security controls, or exposing sensitive information should require approval, he said.

"The goal is not to remove humans from SRE; it is to remove humans from the repetitive operational work that machines can handle safely," he said.

Companies must also decide which systems an AI agent can access, what changes it can make, and when it must ask permission. They need records of everything the agent did — especially when a fix fails.

Cost is another constraint. Applications generate enormous volumes of logs, metrics and operational data, and companies pay to collect, store and analyze it, often across several platforms. AI can add expense if it means shipping large data volumes to external services. Yaari argues more analysis should run closer to where applications live. DataAgent deploys agents within customers' infrastructure to cut data transfer. Its documentation describes local monitoring, cloud-based AI analysis and guided investigation; Yaari describes broader autonomous remediation. The company has not provided independent performance data on how much operational work or cost those capabilities eliminate.

What happens to the engineer?

Yaari expects engineers to spend less time investigating repetitive incidents and more time designing systems, setting policies and deciding what automated tools may do.

"The fundamental shift is from people operating infrastructure to people managing autonomous infrastructure," he said.

Smaller teams could support more applications — provided they understand how automated systems make decisions, and what happens when those decisions go wrong. System architecture, security, reliability and investigating unusual failures could grow in importance as routine operations automate.

Google's DORA research found organizations with strong engineering practices were best positioned to benefit from AI-assisted development. The same logic likely applies to AI-powered operations: companies will still need reliable processes for testing changes, responding to incidents and verifying that automated fixes actually work.

The industry has solved speed in writing code. Whether it can solve speed in keeping code alive — with 44% of organizations still demanding a human sign-off — will decide how much of the engineer's job actually changes.

Original: cloud.google.com

Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,178 articles

Related articles

« Previous articleNext article »