> Ultimately, we believe that incorporating Jev-driven diagnosis into an SRE agent’s workflow is a significant step toward effectively combining System One and System Two models.
Why? What’s the end goal? Lower cost? Faster analysis?
I configured an agent to respond to pages via Slack. It has access to ClickStack (for telemetry), Kubernetes, and GitHub. It runs one of the Sonnet models. The cost is so low, responding to less than a dozen pages a day (mostly from a very sensitive error count alarm) that swapping got Jev makes zero sense. This is especially true if the results are less trustworthy.
Hello! One of the blog's co-authors here. Faster analysis and lower cost is the goal. We think that a Jev-driven pipeline can be a quick fast-path filter for a lot of alerts that are recurring or even false-positives.
We can also add a confidence output to this pipeline, so we invoke a heavier reasoning agent (like Sonnet) when something suspicious deserves a closer look.
It seems to me that Jev is designed for when many quick decisions, with low input context, need to be made. SRE work is probably best suited for very few critical decisions that need to be made, with high input context.
That's a good point. At the same time, a lot of an SRE's daily work is also processing recurring alerts that are of less significance. We believe that a Jev-driven pipeline can be a helpful to SREs in quickly analyzing them.
I'm guessing you're getting downvoted due to dismissive tone, but I think there's truth to this.
Labeling something as "don't escalate" that should have been isn't great. Paying 5x and having to wait 10s instead of 100ms or whatever for a reasoning model (with tools?) is very likely worth it.
We want something fast so that it can help the human SRE. Jev won't be deciding what to escalate or de-escalate anyways. Preferably, both the human and LLM will be pinged at the same time and work in parallel.
Yes and with an incredibly large context and incredibly large domain specific knowledge.
We had an LLM based SRE product from a vendor I won't mention because we had an NDA as our management is sucking them off by trading whitepapers for discounts. It was like a drunk monkey with a wrecking ball. Think we had to pull it in under 2 weeks because it took out multiple production systems and lead to an entire cluster failover.
My experience is that SRE-esque work you experience the full range of model, from "wow, it would take us ages to even get to this theory for why something failed" to "well the priority field in this protocol goes from lowest=most important, but LLM wrote code as if it was highest=most important, so if someone actually pushed it to production the company wouldn't have working internet access any more".
And a lot of "wishing up a feature", I was using it to explore solutions for a given problem in too I didn't knew 100% and it pretty much came to same solution I wanted to do but... the capabilities were not there in the tool so it just started making up probable config clauses, and of course, it didn't work.
Even on simpler stuff there were traps, for example in middle of debug session I asked it to modify Gitlab config to add request duration logging, so it added correct config format to a flag that didn't exist (option was there, just under different name), because it didn't bother to read the docs (since then I generally link it the docs first so it doesn't try to remember and get it wrong).
All of that is both very dangerous, and also easily fixed by just having competent operator there. And as a tool it's great, as replacement it is just AI bros delusion
This is an interesting experiment. Curious why you used LLM-as-a-Judge (gpt-6-astra). Did we have a ground truth of actual RCA done by a human to compare against?
Yes! All faults in SREGym-lite (and the parent benchmark SREGym) have their ground truths reviewed by us to ensure that they are accurate to the incident. In addition, we also conducted an analysis to ensure the LLM judge agrees with the human annotators too, which can be found on page 6 here: https://arxiv.org/abs/2605.07161.
I switched to a decision model for my local "AI" voice control thingy. faster-whisper produces the text then onnx scores all my HA entities to determine which one I'm talking about .. then it lists all its actions and a second scoring round to determine what action I want. some regex to extract numbers if I'm saying things like "water the lawn for 10 minutes". It's pretty dumb when compared to what an LLM can do but considering you now have semantic scoring "at your finger tips" without resorting to levenshteins or other crude algorithms like that it's insane how smart it can look.
And to answer the obvious question: because latency. This thing turns on the light or water valve in under 100ms on a laptop from 2001 with 8gb ram .. all running locally on the laptop.
under 100ms on a 2001 laptop answers the why-not-llm question. what happens when the top two entities score close together on a short command like lights off, do you just take the first or ask back?
This is amazing! We are also fascinated by how fast Jev is. It's very much a breath of fresh air in the age when LLMs tend to take a much longer time and reason about tasks.
> SRE work is probably best suited for very few critical decisions
Yeah, and the article puts the cost of all 105 diagnoses at about $0.15 in Jev calls. At a dozen pages a day, I wouldn't worry much about the bill for either model. I'd be more interested in the pass rate: 76.2% vs 77.8% for GPT-5.6 Sol (medium).
I can see trying a cheaper model first if you're handling lots of requests and it can resolve most of them without escalating. A dozen pages a day doesn't seem like a reason to add that extra step.
Yeah, if the volume is small, then cost would not matter as much. But like I mentioned above, faster analysis and lower cost (than a full-fledged LLM!) is the goal, especially considering how much faster Jev is than an LLM. A Jev-driven pipeline can be a quick fast-path filter for a lot of alerts that are recurring or even false-positives.
Using Jev for SRE seems like such a backwards prospect. You don't need split-second velocity, you don't need costs savings. What you DO need and Jev provides not at all - is auditability. You want to know WHY and HOW the calls were made, as with humans so with AI. Adding a black box in there seems incredibly misguided, no?
The same way I'd not trust an SRE when they've "got a feeling", I don't intend to start trusting Jev. The audit of process doesn't allow audit of reasoning.
I had the same feeling. it seems like jev is a classifier that's smarter than bag of words but I would not think it would do well to ask it to judge code fix effectiveness.
Auditability does not come from an LLM or Jev but from the human operator themselves. Our Jev-driven diagnosis pipeline here is in the same role as all the AI SRE technologies on the market: they are tools to the human operator. If the SRE simply trusts all the AI tools, then I also wouldn't trust them just as when "they've 'got a feeling'".
clintonb | 19 hours ago
Why? What’s the end goal? Lower cost? Faster analysis?
I configured an agent to respond to pages via Slack. It has access to ClickStack (for telemetry), Kubernetes, and GitHub. It runs one of the Sonnet models. The cost is so low, responding to less than a dozen pages a day (mostly from a very sensitive error count alarm) that swapping got Jev makes zero sense. This is especially true if the results are less trustworthy.
victor9000 | 14 hours ago
s_Hogg | 12 hours ago
cyanydeez | 11 hours ago
yimingsu | 7 hours ago
We can also add a confidence output to this pipeline, so we invoke a heavier reasoning agent (like Sonnet) when something suspicious deserves a closer look.
beebmam | 19 hours ago
yimingsu | 6 hours ago
codeduck | 4 hours ago
An SRE goes and fixes the cause of the recurring alerts.
N_Lens | 17 hours ago
yehosef | 12 hours ago
soltanov | 17 hours ago
jasonjmcghee | 16 hours ago
Labeling something as "don't escalate" that should have been isn't great. Paying 5x and having to wait 10s instead of 100ms or whatever for a reasoning model (with tools?) is very likely worth it.
yimingsu | 4 hours ago
sdcfgy | 14 hours ago
We had an LLM based SRE product from a vendor I won't mention because we had an NDA as our management is sucking them off by trading whitepapers for discounts. It was like a drunk monkey with a wrecking ball. Think we had to pull it in under 2 weeks because it took out multiple production systems and lead to an entire cluster failover.
The meat sacks now know they have job security.
PunchyHamster | 13 hours ago
And a lot of "wishing up a feature", I was using it to explore solutions for a given problem in too I didn't knew 100% and it pretty much came to same solution I wanted to do but... the capabilities were not there in the tool so it just started making up probable config clauses, and of course, it didn't work.
Even on simpler stuff there were traps, for example in middle of debug session I asked it to modify Gitlab config to add request duration logging, so it added correct config format to a flag that didn't exist (option was there, just under different name), because it didn't bother to read the docs (since then I generally link it the docs first so it doesn't try to remember and get it wrong).
All of that is both very dangerous, and also easily fixed by just having competent operator there. And as a tool it's great, as replacement it is just AI bros delusion
endangeredhuman | 16 hours ago
yimingsu | 7 hours ago
amne | 15 hours ago
And to answer the obvious question: because latency. This thing turns on the light or water valve in under 100ms on a laptop from 2001 with 8gb ram .. all running locally on the laptop.
yehosef | 12 hours ago
nikolovv | 9 hours ago
yimingsu | 7 hours ago
olgava | 15 hours ago
Yeah, and the article puts the cost of all 105 diagnoses at about $0.15 in Jev calls. At a dozen pages a day, I wouldn't worry much about the bill for either model. I'd be more interested in the pass rate: 76.2% vs 77.8% for GPT-5.6 Sol (medium).
I can see trying a cheaper model first if you're handling lots of requests and it can resolve most of them without escalating. A dozen pages a day doesn't seem like a reason to add that extra step.
yimingsu | 6 hours ago
tinthedev | 13 hours ago
The same way I'd not trust an SRE when they've "got a feeling", I don't intend to start trusting Jev. The audit of process doesn't allow audit of reasoning.
yehosef | 12 hours ago
yimingsu | 4 hours ago
codeduck | 4 hours ago