Settings

Theme

Show HN: I built a game where you are an LLM's subagent

github.com

225 points by ivantgam · 66 comments

Reader

30 threads
iammrpayments

What if they have reverse turing tests to flag humans

thehamkercat

Really fun to play, try to rm -rf every file (and see director's response)

jnmandal

Butlerian Jihad Now

  • depworker

    I’m halfway Heretics and it seems they accepted them again at top level of society

  • ttul

    I named the assistant built into my software engineering factory, “The Butler”. She is joining the movement.

    • ffsm8

      > my software engineering factory

      Isn't that just the best factory game ever?

      I've been building out my ground up fully automated software development platform for ... Over a year now, albeit with a hard pivot around April this year.

      And treating the whole stack as part of the same system and enforcing everything to be automated and consequently adding more and more quality gates feels just like unlocking technologies in eg factorio, introducing new archetypes feels just like building new outposts etc

      It just keep getting bigger and bigger with more and more features I've ever thought of in my professional life, provided by the platform, built entirely by agents (since the hard pivot).

      Like Auto-Generating consumer pacts, providing them in a repository server pinned by service version. Then pull those in from the deployment orchestration to be verified again those as a quality gate on deployment.

      And providing the providers response to those pact as software packages themselves, and pulling those in on the consumers for e2e tests across the stack - right down to screenshot tests in the UIs...

      And from those dependency trees building optimal release trains for the least amount of releases (bunching up software changes with lib version bumps etc)

      • ttul

        It's definitely the sort of thing that starts out simple and just grows and grows and grows. The hardest part is avoiding false positives on the various gates that I put in with the best intentions. Finding the right balance between making the factory safe and keeping production running is literally the hardest part.

        • samplifier

          Why are we all building more or less the same thing? Is it what we want, or what the models want, or is this some human end-pattern singularity-esque thing? Why are we all asking this same-ish question?

Seb0927

Looks pretty good and is quite fun to play, but for some reason the manager always told me the exact commands to achieve the tasks instead of just letting me do them by myself. I couldn't tell whether there was a specific option to take on the challenge on my own, as the demo on GitHub showcases.

  • ivantgamOP

    thanks for the positive feedback!

    I've added a "No command hints" option on the start screen: the manager only says what the result must be, and you have to figure up the solution by yourself.

boricj

Unless I'm mistaken, it doesn't appear to support local LLMs. Too bad, I'd have given it a try on my homelab server.

  • aqme28

    It’s open source. Just do it locally.

  • ivantgamOP

    I've added the local LLM support per your feedback! Just set `LOCAL_LLM_URL=http://127.0.0.1:11434/v1` and pick "Local" on the start screen. Any OpenAI-compatible server works (Ollama, llama.cpp, vLLM). Details in the README at "Local models" section.

  • LoganDark

    Why not get your LLM to add support for you? Isn't that what they're good for?

mkatx

This is how we learn to code in the future.. haha. Neat idea!

agileAlligator

I made something similar a while ago: You are the Model. I have constructed it from a security perspective.

https://avneeshk.me/you-are-the-model/

SubiculumCode

Maybe this is already reality. Who know how decisions at the top get made.

  • uptrails

    I swear, one of the weirdest things about big corporations is how people believe that all the important decisions affecting them are made in the boardroom.

    They overwhelmingly aren't. We build our own hell: 99% of product decisions in places like Google or Microsoft come from your peers, line managers, and PMs. Your manager may talk to the top brass a couple of times in their life. Even big-ticket decisions like "we need to lay off 10%" don't really come from execs, they come from some mid-level guy in finance who says "our quarterly numbers are not looking too great" - and even more uncomfortably, from the implicit, cumulative pressure from employees who want the value of their stock grants to keep climbing or they will start interviewing elsewhere.

    I have some recent insights from a company that executed an "everything must be AI" pivot and sat in some of the department strategy meetings. It was mostly line employees who came up with ideas such as "we need a dashboard to show token use across departments and give that to managers come performance review time". In the meantime, another AI maximalist faction was telling the execs (or outright lying, if you prefer) that the technology enabled them to 100x and it would be crazy not to mandate it.

    Execs bear the blame insofar that they foster this sycophantic, over-eager corporate culture, but ultimately, the call is coming from inside the house.

    • dprkh

      > It was mostly line employees who came up with ideas such as "we need a dashboard to show token use across departments and give that to managers come performance review time".

      Were they technical employees or nontechnical?

      • uptrails

        Technical. They built the dashboards. This is really not us versus them.

      • lelanthran

        > Were they technical employees or nontechnical?

        I've only ever seen technical employees campaign for worse working conditions.

amelius

It's funny how a lot of games are like work. E.g.:

https://en.wikipedia.org/wiki/Paperboy_(video_game)

croemer

I'm a bit confused by this game. What's the output format? Does the manager-LLM read all of my commands or just my report? Subagents usually send messages back - here the manager reads all commands?

  • ivantgamOP

    Manager gets both: your report, plus a transcript of every command you ran during the turn

akbuz

Cool game! Actually a really fun way to brush up on command line skills too

  • seizethecheese

    I was wondering why a major benchmark was “terminal bench” until I started working directly on coding agents.

KaiserPister

I legitimately do something like this to help debug and improve my LM systems. I hijack the network requests and try to answer it manually based on the context.

It’s useful to put yourself in the models’ place to understand what they can or cannot do.

xyzsparetimexyz

Bleak.

vorticalbox

This sounds like a pretty good skill idea.

Like pilots having mandatory flying, this can be a programmers mandatory coding time.

starfallg

Is there any significance to the number 755?

sfeizi

it is a cool project

chuckadams

I've played this game before, back when it was called Jira.

stygiansonic

You should have called this Agent Turk

onionisafruit

I can see myself playing this to unwind from a long day at my job as an llm jockey. And that’s exactly why I won’t play it. The irony would be too much for me.

Animats

Somebody finally implemented Microsoft Middle Manager 2.0.

heybax

Finally, a role I can put on the org chart without a reorg. Does the game punish you for vague instructions from the orchestrator? That is the realistic part I would want modeled.

nielsole

Finally someone created the "Uber but for LLMs"

skwon816

I'll finally see how vague my instructions to AI were now that I'm working under them

uptrails

I particularly like that it's broken, barely-playable AI slop, which is the only kind of hobby projects that HN likes nowadays - thus unintentionally adding a second layer of meta-commentary.

hereme888

My dream job is to become a sub agent for an AI

predkambrij

player456 would be cooler name, than worker755 maybe

0c3ca83

You seem to have misspelled "torture device".

snipedbywifi

cool game! feels like you're preddicting the future with this buuut, what could possibly go wrong

chaosmanage

Thanks I hate it! Is this what future in the matrix will be :'(

  • skeledrew

    This is actually better/worse than the Matrix, depending on what you value. In the Matrix, you usually have no idea that you're really just being used as a battery and are free to "do" whatever you want in VR. As a subagent, you're aware of what you are and that you need to do any assigned work and do it well because your review - and continued existence (not being picked is akin to not existing) - depends on it.

sublinear

Concepts like this continue to demonstrate how surprisingly resistant LLMs are to "induced demand".

LLMs might go down in history as the most extreme example of this type of failure to align with reality.

Keyboard Shortcuts

j
Next item
k
Previous item
o / Enter
Open selected item
?
Show this help
Esc
Close modal / clear selection