Building better AI tools
hazelweakly.me343 points by eternalreturn 3 days ago
343 points by eternalreturn 3 days ago
This is a confusing piece. A lot of it would make sense if Weakly was talking about a coding agent (a particular flavor of agent that worked more like how antirez just said he prefers coding with AI in 2025 --- more manual, more advisory, less do-ing). But she's not: she's talking about agents that assist in investigating and resolving operations incidents.
The fulcrum of Weakly's argument is that agents should stay in their lane, offering helpful Clippy-like suggestions and letting humans drive. But what exactly is the value in having humans grovel through logs to isolate anomalies and create hypotheses for incidents? AI tools are fundamentally better at this task than humans are, for the same reason that computers are better at playing chess.
What Weakly seems to be doing is laying out a bright line between advising engineers and actually performing actions --- any kind of action, other than suggestions (and only those suggestions the human driver would want, and wouldn't prefer to learn and upskill on their own). That's not the right line. There are actions AI tools shouldn't perform autonomously (I certainly wouldn't let one run a Terraform apply), but there are plenty of actions where it doesn't make sense to stop them.
The purpose of incident resolution is to resolve incidents.
There's no AI tool today that will resolve incidents to anyone's satisfaction. People need to be in the loop not only to take responsibility but to make sure the right actions are performed.
Exactly. There seems to be this fantasy in which you can somehow string different kinds of agents together, one designing and one reviewing, and that finally producing something superior as output - I just don't buy that.
Sounds like heuristics added on top of statistics, which is trying to remedy some root problem with another hack.
The whole field of metaheuristic algorithms rests on a similar idea. a lot of stupid "agents" finding a good solution. by metaheuristics i mean genetic algorithms, PSO, ACO etc.
Hmm, but this provably works right now though? All LLMs perform better with roleplay direction and focused scope. Using coding agents with plan then execute makes noticeable quality improvements.
Why isn't this the de facto then? Anyone packaging such commercial solutions?
Nobody disputes this. Weakly posits a bright line between agents suggesting active steps and agents actually performing active steps. The problem is that during incident investigations, some active steps make a lot of sense for agents to perform, and others don't; the line isn't where she seems to claim it is.
Understood. To your example about the logs, my concern would be be that the AI chooses the wrong thing to focus on and people decide there’s nothing of interest in the logs, thus overlooking a vital clue.
You wouldn't anticipate using AI tools to one-shot complex incidents, just to rapidly surface competing hypotheses.
It's not a confusing piece if you don't skip/ignore the first part. You're using her one example and removing the portion about how human beings learn and how AI is actively removing that process. The incident resolution is an example of her general point.
I feel pretty comfortable with how my comment captures the context of the whole piece, which of course I did read. Again: what's weird about this is that the first part would be pretty coherent and defensible if applied to coding agents (some people will want to work the way she spells out, especially earlier in their career, some people won't), but doesn't make as much sense for the example she uses for the remaining 2/3rds of the piece.
It makes perfect sense for that case too. If you let AI do the whole job of incident handling (and leaving aside the problem where they'll get it horribly wrong), that also has the same problem of breaking the processes by which people learn. (You could make the classic "calculator" vs "long division" argument here, but one difference is, calculators are reliable.)
Also:
> some people will want to work the way she spells out, especially earlier in their career
If you're going to be insulting by implying that only newbies should be cautious about AI preventing them from learning, be explicit about it.
You can simply disagree with me and we can hash it out. The "early career" thing is something Weakly herself has called out.
I disagree with you that incident responders learn best by e.g. groveling through OpenSearch clusters themselves. In fact, I think the opposite thing is true: LLM agents do interesting things that humans don't think to do, and also can put more hypotheses on the table for incident responders to consider, faster, rather than the ordinary process of rabbitholing serially down individual hypothesis, 20-30 minutes at a time, never seeing the forest for the trees.
I think the same thing is probably true of things like "dumping complicated iproute2 routing table configurations" or "inspecting current DNS state". I know it to be the case for LVM2 debugging†!
Note that these are all active investigation steps, that involve the LLM agent actually doing stuff, but none of it is plausibly destructive.
† Albeit tediously, with me shuttling things to and from an LLM rather than an agent doing things; this sucks, but we haven't solved the security issues yet.
The only mention I see of early-career coming up in the article is "matches how I would teach an early career engineer the process of managing an incident". That isn't a claim that only early career engineers learn this way or benefit from working in this style. Your comment implied that the primary people who might want to work in the way proposed in this article are those early in their career. I would, indeed, disagree with that.
Consider, by way of example, the classic problem of teaching someone to find information. If someone asks "how do I X" and you answer "by doing Y", they have learned one thing (and will hopefully retain it). If someone asks "how do I X" and you answer "here's the search I did to find the answer of Y", they have now learned two things, and one of them reinforces a critical skill they should be using throughout their career.
I am not suggesting that incident response should be done entirely by hand, or that there's zero place for AI. AI is somewhat good at, for instance, looking at a huge amount of information at once and pointing towards things that might warrant a closer look. I'm nonetheless agreeing with the point that the human should be in the loop to a large degree.
That also partly addresses the fundamental security problems of letting AI run commands in production, though in practice I do think it likely that people will run commands presented to them without careful checking.
> none of it is plausibly destructive
In theory, you could have a safelist of ways to gather information non-destructively. In practice, it would not surprise me at all if pople don't. I think it's very likely that many people will deploy AI tools in production and not solve any of the security issues, and incidents will result.
I am all for the concept of having a giant dashboard that collects and presents any non-destructive information rapidly. That tool is useful for a human, too. (Along with presenting the commands that were used to obtain that information.)
Previous writing, Josh, and I'm done now litigating whether I wrote the "early career" thing in bad faith and expect you to be too.
I don't see you materially disagreeing with me about anything. I read Weakly to be saying that AI incident response tools --- the main focus of her piece --- should operate with hands tied behind their back, delegating nondestructive active investigation steps back to human hands in order to create opportunities for learning. I think that's a bad line to draw. In fact, I think it's unlikely to help people learn --- seeing the results of investigative steps all lined up next to each other and synthesized is a powerful way to learn those techniques for yourself.
I’m going to but in here.
I think the point the article is making is to observe the patterns humans (hopefully good ones) follow to resolve issues and build paths to make that quicker.
So at first the AI does almost nothing, it observes that in general the human will search for specific logs. If it observes that behaviour enough it then, on its own or through a ticket, builds a Ui flow that enables that behaviour. So now it doesn’t search the log but offers a button to search the log with some prefilled params.
The human likely wanted to perform that action and it has now become easier.
This reinforces good behaviour if you don’t know the steps usually followed and doesn’t pigeonhole someone into an action plan if it is unrelated.
Is this much much harder, yes it is than just building an agent that does X. But it’s a significantly better tool because it doesn’t have humans lose the ability to reason about the process. It just makes them more efficient.
We're just disagreeing and hashing this out, but, no, I don't think that's accurate. AI tools don't watch what human operators in a specific infrastructure do and then try to replicate them. They do things autonomously based on their own voluminous training information, and those things include lots of steps that humans are unlikely to take, and that are useful.
One intuitive way to think about this is that any human operator is prepared to bring a subset of investigative approaches to bear on a problem; they've had exposure to a tiny subset of all problems. Meanwhile, agents have exposure to a vast corpus of diagnostic case studies.
Further, agents can quickly operate on higher-order information: a human attempting to run down an anomaly first has to think about where to look for the anomaly, and then decide to do further investigation based on it. An AI agent can issue tool calls in parallel and quickly digest lots of information, spotting anomalies without any real intentionality or deliberation, which then get fed back into context where they're reasoned about naturally as if they were axioms available at the beginning of the incident.
As a simple example: you've got a corrupted DeviceMapper volume somewhere, you're on the host with it, all you know is you're seeing dmesg errors about it; you just dump a bunch of lvs/dmsetup output into a chat window. 5-10 seconds later the LLM is cross referencing lines and noticing block sizes aren't matching up. It just automatically (though lossily) spots stuff like this, in ways humans can't.
It's important to keep perspective: the value add here is that AI tools can quickly, by taking active diagnostic steps, surface several hypotheses about the cause of an incident. I'm not claiming they one-shot incidents, or that their hypotheses all tend to be good. Rather, it's just that if you're a skilled operator, having a menu of instantly generated hypotheses to start from, diligently documented, is well worth whatever the token cost is to generate it.