Details
"Vibe coding" has revolutionized the IDE, making development fast, fluid, and improvisational. Naturally, engineers are trying to bring that same magic to incident response—a trend Gartner now calls Vibe Ops. But the moment you point a raw, unguided LLM at production operations, the vibe turns into a "vibe oops."
In this sharp, 10-minute lightning talk, NeuBird Field CTO Francois Martel breaks down why production environments completely invert the rules that make AI coding safe. While coding offers cheap verification and easy rollbacks, production offers neither. You will learn why pasting raw log dumps into LLMs creates a garbage-in, garbage-out cycle of costly token waste and authoritative hallucinations. Furthermore, Francois explores how unmanaged AI usage fractures incident response into isolated, unversioned investigations, creating an audibility and security nightmare.
Perfect for SREs looking to ground their AI toolset, and tech executives looking to eliminate the hidden compliance and downtime risks of "shadow AI" in production. Attendees will leave this session understanding:
The Production Inversion: Why AI tools require radically different architectures for operations than they do for software development.
The Context Problem: Why the secret to AI troubleshooting isn't a better prompt, but rather intelligent telemetry refinement.
The Path Forward: How to transition from risky, siloed AI experimentation to a collaborative, governed system that builds institutional memory and solves incidents safely.
Improvisation is Dangerous in Production Environments: Relying on unguided AI "vibing" or starting incident investigations from scratch is highly inefficient and risky given the sheer volume of data and alerts.
The Necessity of a Structured Context Layer: High-accuracy AI troubleshooting requires a reliable layer that maps raw production infrastructure connections into automated dependency graphs, aiding functions like blast radius detection.
Automated Closed-Loop Remediation: Modern operations can link production monitoring agents with development coding agents to automate the lifecycle of an incident—from root-cause identification to opening Jira tickets and merging pull requests.
Citation-Backed and Data-Driven RCA: To maintain compliance and auditability, AI agents must quote their sources and verify telemetry by executing code rather than fabricating conclusions.
Key Insight: Raw production environments contain too much data and missed signals for manual, unassisted triage.
Practical Implication: Organizations must shift away from manual telemetry deep dives and deploy specialized context-aware AI agents.
Key Insight: Ad-hoc prompting in developer environments consumes significant time and excessive cloud token costs.
Practical Implication: Standardizing around background AI agents executing targeted tools streamlines root cause analysis to a matter of minutes.
Key Insight: Proactive scheduled scans can evaluate risk assessments and health checks before a hard failure occurs.
Practical Implication: Teams can automate regular posture reports and seamlessly flag high-risk infrastructure straight into ticketing systems like Jira.
Q: What is the main issue with using standard, unguided AI during a production incident?
A: Production features a massive amount of complex telemetry and alerts. Standard troubleshooting without integrated structural context leads to excessive token consumption, slower resolution times, and overreliance on improvisational troubleshooting.
Q: How does the dependency graph improve incident analysis?
A: It automatically maps raw environment connections to expose how services link to foundational infrastructure, making it fast and simple to isolate the blast radius of an alert.
Q: Can these AI operations agents collaborate directly with development tools?
A: Yes. When an incident is discovered, the operations agent can automatically log a Jira ticket or GitHub issue. In contrast, a linked coding agent generates and submits a pull request to resolve the software flaw.
Hello everyone and welcome to this 10-minute take. My name is Frano Martell. I'm Field CTO at Neubird AI, and today we're going to talk about Vibe Ops and why AI improvisation fails in production.
The basic problem is that there's a lot of data and a lot of alerts, and still, you're missing some alerts. So, people have a hard time dealing with data in production. There's too much of it, and the last thing you want to do is improvise and start from scratch every time. So, what we've built is an agent that has context coming from your production environment. All these connections at the bottom, that context is enriched and organized into tools that make investigations much faster for an agent. Okay. And these agents have skills and, basically, a structure that ensures accuracy and cost efficiency. Then, you can interact with this platform through Slack, Teams, through this Neubird UI—which I'm going to show you in a minute—or in your favorite vibe coding environment, like Cloud Code, for example, can connect to our agent and delegate the hard work of understanding why things are failing in production.
This is the web user interface, and what you're seeing here is a view of our dependency graph. This is one of the things that we built automatically over time from the raw connections. You can imagine having a dependency graph showing which service is connected to what infrastructure—it really helps doing things like blast radius detection. Let me increase the font here so you can see a little bit. You know, I've got alerts here that have been triaged, and I've got the ability to get details on those and run investigations. The investigations can run automatically, and they can be triggered manually.
I'm seeing here my history of investigations. You probably caught a glimpse of the confidence score. A 99% confidence score is industry-leading, and it's because we have a very data-driven root cause analysis. What you'll see here in the report, the RCA report, is a very quick view that you can act on immediately. You have the root cause analysis summary here, and if you scroll down, I can go straight to the remediation plan and take action.
In this case, I had an alert, so I can go show you the signal context, right? What triggered this? It was a latency P95 alert on my order service, and the root cause was an Oracle TNS listener failure on port 1521. Immediately, I can go and see exactly what to do about it. Imagine this in comparison to hearing that you have some latency problem, opening up your cloud code, and starting. If you have done it right, you'll have your connections already set up and you'll start looking for things. By the time you get to these results, you'll have spent a lot of tokens, and it will take way more than the five minutes that this investigation took, right? This happens all in the background while nobody's actually looking. They get this information, maybe on a Slack channel; they'll get a link to this investigation. You'll see that we quote all our sources, and we're actually running code in some cases to get to the data, so there's always citations showing where the information is coming from. Okay. So, this is an example of a reactive investigation.
What I'm going to show you next is an example of a proactive investigation. I'm going to switch to this screen here and show you. We have these agent tasks that can be run, okay. Your agent tasks are prompts, so the product can be prompted at any time. I can go to the console and run a prompt. By the way, as part of an investigation, I can ask follow-up questions and dig deeper. I can, you know, help the agent learn as part of those interactions. But these prompts here can get scheduled, and they could be tied to a workflow. I've got some examples here: you might want to do a change risk assessment, or you want to do an observability posture report. It'll have different cadences, and then you'll get reports for these.
What I want to show you is this investigation where we're doing a scan for HubSpot form submission issues, and we're running those on a schedule. We have all these recent runs. You can see the report here with the root cause analysis, the same kind of format you saw earlier. In this case, it's tied to a code repo, so we have a kind of coding change suggestion.
The workflow that's tied to this actually went and created a Jira ticket, a GitHub issue, and a GitHub pull request. If I share the Jira ticket here, you'll see that the root cause analysis was created in Jira. Then, if I show you the pull request, it's actually a coding agent that created the code change to fix the problem based on the GitHub issue we opened. So, we can collaborate with your coding agents.
You can have this closed-loop system where you're writing code very quickly with your agents, deploying that code, having our agent for production operations monitor the health of your applications, and then handling issues automatically by going back to your coding agents to fix those issues. You can do this fully automated end-to-end and get your velocity of problem-fixing up significantly.
It all is founded on this idea that you need high accuracy. To get high accuracy, you need a really solid context layer that sits between your data and your agent so that the agent can discover dependencies, blast radius, do alert filtering, and really streamline the work that's required and the tokens required in order to arrive at a root cause analysis or any sort of reports. Right? I have a health check report that I can generate, and I've got all these different investigations. Like this one is a risk assessment report where it's giving me the high-risk items, the medium-risk items, and the low-risk items. With the workflow feature, I can create Jira tickets on my high-risk items automatically, for example.
Hopefully, this kind of presentation helps you wrap your head around what's possible in production, where vibing is dangerous. I don't advise it. I think you need something that's very structured, that's auditable, that is going to be more predictable and cost-efficient, and that also learns across every interaction across the full enterprise landscape.
Thank you for your time today. Hopefully, this will motivate you to come and see what we can do at neubird.ai. neubird.ai is where you can go and find out more information, ask for a demo, or see it in action.
Presenters:
Francois Martel
Field CTO
Join Now for More Content & Events
For event and sponsorship inquiries, please email: [email protected]
