agent's failures on long horizon tasks

We've moved from LLMs being able to work on a task for about 2 minutes to about 2 hours in the last 18 months, and that's mostly limited by the context window size filling up. In 14 years time I don't really see a reason why that wouldn't have extended a time frame that's effectively continuous forever, or at least a ceiling that's indistinguishable from that.

The question really becomes "why would we want that?". The main reason you'd want an AI that can focus on a task forever is to completely remove the human from the loop. That's something we should be cautious about.

▲ FinnKuhn a day ago | parent | next [-]

  I do not think agents need to work continuously. They "just"
  need to work on a project longer than an employee is able to
  work on it for them to be commercially useful. Although this
  does not consider that agents might take a shorter or longer
  amount of time for the same task. Therefore, we should begin
  to compare these in tasks completed during that time instead.

	▲	onion2k a day ago \| parent [-]
		Not sure about that. The nature of work changes if you have some[one\|thing] that can work on it continuously forever. The goal of using AI shouldn't be to do the things a person does; it should be to do the things a person would be able to do if they had (practically) unlimited time. This is one of the inflection points around working with AI. When your thinking shifts from "AI does what the person used to do" to "AI does something different that leads to the same outcome as before, but with all that cool stuff I'd love to have the time for", then the equation changes. For example, I will never write another app that doesn't have 100% test coverage again. I stress- and soak-test everything these days. That's changing how I write code - I need to write things in ways that have deterministic harnesses for things mutable state outside of my immediate control (like random numbers or datetimes or state loaded from a save) so that I can task AI with building a fuzzer that does tens of thousands of random tests on every push to see if anything broke. I couldn't do that in the past because it was always effort that a client wouldn't get enough value from to pay me to do it. I can do those things now. It's ace. (Everywhere I put 'I' you can replace with 'AI of the Week')

▲ bbmatryoshka a day ago | parent | prev | next [-]

2 hours... of agent work, usually for the same output the average worker has to use much more time

	▲	knollimar a day ago \| parent [-]
		Throw understanding images in there and it becomes less true in my experience. Sure they shit out code and search text well, but one infographic and they circle trying to rasterize it.

▲ Gareth321 a day ago | parent | prev [-]

My codex tasks regularly cross 8 hours, and I'm only using Sol High. It's not unusual for tasks to span much longer. It just requires instructions to continue working until the spec is complete.

Anthropic and OpenAI are currently obsessed with getting humans out of the training and improvement loops. It's going to happen very soon and when it does I think we see staggering improvements in a very short space of time. Basically, the Singularity.

▲

munksbeer 21 hours ago | parent | next [-]

Are you breaking up your tasks and spawning new sessions for each, or are you just yoloing and letting it auto compact when it blows the context many times on such a long running task?

	▲	Gareth321 21 hours ago \| parent [-]
		Yolo. I keep a large project markdown document + incident/logs/feature documents which its instructed to review and update as necessary. It still occasionally misses stuff but it's surprisingly effective. Disclaimer: this is for hobby software. For work I'm more cautious - usually.

▲

bwfan123 a day ago | parent | prev [-]

> It just requires instructions to continue working until the spec is complete

Try putting an LLM agent in a deterministic workflow without humans in the loop. My experience with this is not encouraging. Getting it to work requires sprinkling some context magic and hoping and praying the LLM does the right thing. More astrology or religion and less science. Great for use cases with humans-in-the-loop, but less than impressive when you need determinism and reliable operation.

▲

Gareth321 21 hours ago | parent [-]

> Try putting an LLM agent in a deterministic workflow without humans in the loop.

I would not use a sewing machine to repair my deck :) LLMs are non-deterministic, by design.

	▲	onion2k 21 hours ago \| parent [-]
		LLMs are non-deterministic, by design. They don't have to be though. You can give them a temperature of zero so they pick the highest-probability token every time to give you a deterministic output. I imagine this doesn't work on frontier models because there's a lot going on, but you can definitely do it with a small local model.