Ask HN: Examples of agentic LLM systems in production? Now that everybody and their mother are fuzzing in social media about LLM agents and agentic LLM systems (or something), are there actual examples of live applications that are based on an agentic LLM process flow? I'd be curious to know and see such examples in order to derive some inspiration from them. |
issue however, is that the agents cannot be chained. ie, chaining requires deterministic outputs and not narratives.
I think what you mean is that they work best in cases where it's very hard to measure how well they are working.
Actual real Intelligent Autonomous Agents - go to Mars, kick a rover... there's one. Go try front running on the markets, you'll meet about 6000 other ones trying to out run you.
I do contracting work, we're building a text-to-sql automated business analyst. It's quite well-rounded: it tries to recover from errors, allows automatic creation of appropriate visualisations, has a generic "faq" component to help the user understand how to use the tool. The tool is available to some 10.000 b2b users.
It's just a bunch of prompts conditionally slapped together in a call graph.
The client needed AGENTIC AI, without specifying exactly what this meant. I spent two weeks pushing back on it, stating that if you replace the hardcoded call graph with something that has """free will""", accuracy and interpretability goes down whilst runtimes go up... but no, we must have agents.
So I did nothing, and called the current setup "constrained agentic ai". The result: High fives all around, everyone is happy
Make of that what you will... ai agents are at least 90% hype.
I've implement countless LLM based "agentic" workflows over the past year. They are simple. It is a series of prompts that maintain state with a targeted output.
The common association with "a floating R2D2" is not helpful.
They are not magic.
The core elements I'm seeing so far are: the prompt(s), a capacity for passing in context, a structure for defining how to move through the prompts, integrating the context into prompts, bridging the non-deterministic -> deterministic divide and callbacks or what-to-do-next
The closest analogy that I find helpful is lambda functions.
What makes them "feel" more complicated is the non-deterministic bits. But, in the end, it is text going in and text coming out.
Are you using a separate state manager + function calling so the LLM knows where it is?
I love those stories but I could never do that with a straight face. However, the AI field is such an uphill battle against all the crap that LinkedIn influencers are pushing into the minds of the C-suite... I feel it's okay to get a bit creative to get a win-win here ;)
I guess everything with an algorithm in it is AI if you look at it from enough of a distance...
Everything is very boring tech-wise, using vanilla postgres/pgvector and a few hundred lines of python. Every RAG-searchable text field (mostly column descriptions and a list of LLM-generated example queries) is linked to nodes holding metadata, at most 2 hops out. The tool is available to 10.000 users, but load is only a few queries per minute at peak... so performance wise it's fine.
Hope you can find a tool; the big data players are of course jumping on this (snowflake, databricks, they all talk about their text-to-sql tools).
If you have the budget and want something bespoke built that has some magic sauce tuned to your exact problem field, send me an email!
The right approach to build toward agents is to start with something that gives pretty good responses to prompts and build up an agentic mode to let it do more and more in response to each prompt. It should be thought of as extending how much you get per prompt, and doing so by chaining together components you've already worked at making to good at.
Cursor (the LLM powered VS Code fork) has an agentic mode and they are doing this the right way. The normal chat window is good at producing changes to your code, and at applying them, at looking at lints, at suggesting terminal commands, at doing directory listings or RAG on your codebase. Agentic mode is tying those together to do more of the work you want with fewer prompts from you.
1. Its still perceived as an issue of competitive advantage
2. There is a serious concern about backlash. The public's response to finding out that companies have used AI has often not been good (or even reasonable) -- particularly if there was worker replacement related to it.
It's a bit more complicated with "agents" as there are 4 or 5 competing definitions for what that actually means. No one is really sure what an 'agentic' system is right now.
I look at my workplace and I see places where they might fit in but if the reliability isn’t 99.5% they won’t be trusted and I think that’s a problem.
I made a toy in n8n that collects transactions in YNAB via API and matches them to Amazon orders in GMail. It then uses GPT-4o with vision to categorize the product pictures according to my budget’s categories but I have to add the order link to the transaction memo and add a flag for human review because it’s only 80% or so. It has sped up the workflow for sure but nowhere near good enough to set it and forget it.
I had no idea about Make.com or n8n, they seem interesting. Thanks for the tip! Will check them out.
And non ui flow diagram but essentially the same thing: inngest, hatchet.
- Auth bypass/arbitrary file read in Scoold: https://xbow.com/blog/xbow-scoold-vuln/
- SSRF in 2FAuth: https://xbow.com/blog/xbow-2fauth-ssrf/
- Stored XSS in 2FAuth: https://xbow.com/blog/xbow-2fauth-xss/
- Path traversal in Labs.AI EDDI: https://xbow.com/blog/xbow-eddi-path/
Each of those has an associated agent trace so you can go read exactly what the agent did to find and exploit the vulnerability.
If you ask two different people in the AI space to define "agent" you almost always get two slightly (or significantly) different definitions!
Here are just some of the definitions I've seen over time: https://news.ycombinator.com/item?id=42216217#42228364
For the purpose of this thread the most cynical definition, "LLMs that do something useful", might actually be the best fit!
Agentic workflows are a much higher bar that are just barely starting to work. I can't speak to their efficacy but here's a few of the ones that are sort of starter-level agents that I've started seeing some companies adopt:
Last week Wednesday I participated in Anthropic's Model Context Protocol hackathon, and built a system with my team partner Zia to automatically search and find restaurants for your dietary preferences and group size.
It also automatically downloads social media of the restaurant to get a vibe for the place.
There's a video of it in action here: https://www.youtube.com/watch?v=c6vGrfHFyu8
And a Github repo here: https://github.com/zia-r/gotta-eat
In practice, achieving this reliability meant repeatedly:
1. Breaking down complex goals into simpler steps: Composing prompts, tool calls, parsing steps, and branching logic. 2. Debugging failures: Identifying which part of the workflow broke and why. 3. Measuring performance: Assessing changes against real metrics to confirm actual improvement.
We tried some existing observability tools or agent frameworks and they fell short on at least one of these three dimensions. So we built our own: https://github.com/PySpur-Dev/PySpur
1. Graph-based interface: We can lay out an LLM workflow as a node graph. A node can be an LLM call, a function call, a parsing step, or any logic component. The visual structure provides an instant overview, making complex workflows more intuitive. 2. Integrated debugging: When something fails, we can pinpoint the problematic node, tweak it, and re-run it on some test cases right in the UI. 3. Evaluate at the node level: We can assess how node changes affect performance downstream.
We hope it's useful for other LLM developers out there
It seems the community has gotten more negative about agentic approaches since then, and it wasn’t pretty then.
ChatGPT is a good example: you ask for an image, and you receive one; you ask for a web search, and the chatbot provides an answer based on that search.
In both cases, the chatbot has the ability to rewrite your query for that tool and is even able to call the tools multiple times based on the previous result.
There are plenty of RAG-capable LLMs in production, but still few products/UX oriented toward agentic work.
An AI product that can make purchases and API requests to external services like delivery drivers, calendars, etc. is still needed to truly enable these "agents" - which right now are basically read-only domain-specific LLMs.
Your rube goldberg contraption that you put together for your borderline-fradulent pitch deck is NOT an assembly-line nor is it a product anyone's gonna buy. Why?
Because cosine similarity search mathematically sucks a* , large context windows, while better, are nowhere close to being fast and practical ( maybe with a small exception of the generic sounding 1M context summaries you now get from gemini flash 2.0 exp ) . You probably don't have any kind of CI/CD setup, no testing at all zero, no benchmarking of accuracy, you probably can't even get lm_eval installed in the first place so no troubleshooting methodoloy, no formal iteration pipeline, you're not putting out a new model every 2 weeks and iterating upon, and YOU at this point probably can't find your own way to your own fkin toilet seat without Cursor's GPS showing your where it is and then writing a whole factory just to open the toilet seat.
You look at the youtube demos and it's just another investor slop to be sold to other sloppy investors. I even asked on uncle Elons twitter if anyone had a demo of agents doing anything in real life, and after 1/4million views the only thing that even worked AT ALL were spambots and Pliny's agent making a sh*tcoin. https://x.com/nisten/status/1808522547169763448
People cook something at home and immediately get delusional thinking they now have an assembly line that's just going to print money... have you ever actually looked at an industrial pasta-maker machine. Do YOU have the skills to make that? I'm sorry but no ammount of shrooms and microdosed-meth pills is gonna get you that.
Agents do not exist yet, they will sooner or later, but right now they're a concept more along the lines of scammy ledger-backed dbs.
You can always prove me wrong with a real life demonstration of an automated tool doing a complex ammount of steps that you'd normally expect an average-ish worker to do for you on a RELIABLE rate basis. I.E. Doing your taxes like your accountant or 10 year old hopefully does.
One we've been using for meeting notes + action items works quite well https://fireflies.ai
(I'm not trying to be accusatory here, just trying to understand how these opinions spread.)
Will agents still matter once models do a better job paying complete attention to large contexts?
This is the only one that makes sense. People want to conflate it with their random vague conceptions of AGI or ASI or make some kind of vague requirement for a certain level of autonomy, but that doesn't make sense.
An agent is an agent and an autonomous agent is an autonomous agent, but a fully autonomous agent is a fully autonomous agent. An AGI is an AGI but an ASI is an ASI.
Somehow using words and qualifiers to mean different specific things is controversial.
The only thing I will say to complicate it though is if you have a workflow and none of the steps give the system an option to select from more than one tool call, then I would suggest that should be called an LLM workflow and not an agent. Because you removed the agency by not giving it more than one option of action to select from.
I implore you to look into that to see how some people relate it to autonomy or AGI or ASI(wrongly, imo - I think shoehorning OOP and UML diagrams plus limited database like memory/context is not a path to AGI. Clever use of final layers, embeddings and how you store/weight them (and even more interesting combinations) may yield interesting results because we can (buzzword warning) transcend written decoding paradigms - the Human brain clearly does not rely on language).
However what gets marketed today is, as you say, not capable of any real agent autonomy like in academia - they are just self-recursive ChatGPT prompts with additional constraining limits. One day it might be more, but libraries now are all doing that from my eye. And recursion has pros but emphasizes the unreliability con/negative of LLMs.
For me, agentic means that at least at some stage, one model is prompting another model.
Which definition of agents are you using there?
Seems to be the definition they're using. Which is a high bar, in my opinion - but it does illustrate the difficulty current systems will have in meeting an exceptionally high bar of quality (human-grade).
Defining it this way and exploring percentage task failure compared to typical (expert?) human doing the same work is valuable insight, in my opinion.
On the other hand, you can define agents as anything that does tool calling, but then it's trivial to create an agent but still non-trivial to meet expectations of the typical consumer because you aren't observing their failure rates.
Right now they are "read-only" which I would call a persona
[1] (https://github.com/eloquentanalytics/pyeloquent/blob/main/RE...)
Another version of Agents is described in the paper "Agent Hospital: A Simulacrum of Hospital with Evolvable Medical Agents"[0] where instead of having one LLM prompted context, you create many that specialize in different areas, and have them communicate to distribute a task. The effect is that the outcome is higher quality than trying to use a single automated LLM.
Shoot me an email if you want to discuss specifics!