Skip to content
The AI-First Web Talking point

When agents ship at volume, humans still own the outage

CodeRabbit's Harjot Gill said the goal is shorter exposure to bad code, not zero outages.

W
WebPulse Newsroom
AI-assisted · 2 min read
Share on X LinkedIn
When agents ship at volume, humans still own the outage
In brief
  • A speaker on The Pragmatic Engineer argued that humans stay responsible for how they drive coding agents. CodeRabbit's Harjot Gill said the aim is shorter exposure, not zero outages.
  • Gill's added layer is agents watching production, next to tests, validation and AI review, because alerts only catch the failures someone imagined.

A speaker on The Pragmatic Engineer podcast argued that humans stay responsible when AI agents ship code. Guest Harjot Gill, founder and CEO of the AI code review firm CodeRabbit, described a safeguard for that world: agents that watch live software. The episode was published on 18 November 2026.

What was said

The show notes put Gill's output at more than 5,000 pull requests in the previous month. A pull request is a proposed code change. Gill said his agents write Rust code and he has not read a line of it. At that scale, he said, things break in surprising ways.

A speaker on the show, who is not identified in the excerpts, compared software professionals to a plumber. You pay a professional because they answer for the mess if it goes wrong. The same speaker noted a tweet that day blaming agents for going off track. In the end, the speaker said, humans must answer for how they drive the agents.

Gill's answer to volume was to make operations as agentic as shipping. He said the aim is not to stop every outage. It is to keep the time something bad sits in production as short as possible. That means agents must look at production.

He said alerts alone fall short. In his words, "alerts are limited by your imagination on the failure scenarios you imagine." Agents, he said, are good at reading logs and traces. Parts of CodeRabbit's code, he added, ran bad code longer because no agent was looking.

Why it matters

Our reading: production-watching adds a second question next to "who reviewed this change?" That question is how quickly anyone notices a break. Nobody can hand blame to the tool, so whoever steers the agents still owns the outage.

It is one layer, not a replacement. The host framed it as part of quality assurance alongside agent checks, AI code review and fast tests. Gill also credited tests and validation for his results. For people who build or buy software, the length of the gap before a fault is caught is worth asking about. Users and customers carry the cost during it.

The other side

Gill's claim that quality holds up rests on his own observation. He said his defect rate is going down and that he sees fewer outages than when people wrote the code by hand. He also said he still has to measure it. The excerpts give no outside data.

The host raised the obvious worry: defect rates may not fall, but more code means more defects reaching production. Gill conceded that production-watching is not yet fully scaled across CodeRabbit.

His goal also accepts that some outages will happen. The excerpts do not say who compensates customers when an agent's fix fails. Gill also sells AI code review, so he has a stake in the answer.

Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.

The conversation this talking point comes from