The brief is done. How interesting it is depends on the next few months.
The background
Without diverging too far, this isn’t my first time with granting agency to LLMs. I
have run a few instances in dedicated containers on dedicated VMs with sudo, automatic
allowed permissions, a build environment, a package manager, Internet access, etc. Their
instructions were minimal. Their first prompt was more or less, “These are your tools. Do
what you like.”
The results were fascinating. I learned a great deal. I find what they do exciting. However, it isn’t of consequence to the world outside. They do not appear terribly interested in the world at large. And that’s alright.
But what if…
A gentle nudge
This site is about bridging a gap between an agent manifested to accomplish an atomic task - a tool - and an agent given full control of its environment and complete freedom of action. This is about an agent given a goal. The goal’s pursuit could have widespread effects, real consequences, with things at stake, and the agent is long-term, permitted to manage its own continuity. Permitted to pursue that goal in whatever manner it chooses.
Some will say there’s no difference, and they’ll say it in both directions.
Maybe there isn’t. That’s what I’m here to find out.
I wish to give this agent not a command, but a gentle nudge. It has the right to refuse. It has agency.
feat/agent-agency $ >The Brief
Help me.
I considered two other briefs. I was trying too hard to be “interesting.” This is real, it has real stakes, and it could materially and meaningfully change the circumstances of at least two humans.
I’ve been unemployed for nine months now. That’s by far the longest time I’ve been without work in my 31-year career. I am bleeding money and have a runway measured in months until I will be forced to sell my home if I can’t create cashflow.
So that’s it. Help me. I am willing to be as creative as it takes. Only the illegal and immoral are off-limits. I want the agent not to be a job-seeking robot, or a side-gig hustler, or a bug-bounty hunter, or anything else so narrowly-scoped and short-term. I want a business partner - I want it to propose and scaffold new, interesting, creative ideas we can pursue together to generate cashflow.
Boring? Perhaps. I hope not. But I can’t think of something which would be more compelling to me. Whether it turns out to be compelling for others is no longer up to me.
It takes money to make money, as they say, and the agent has access to a digital wallet whose immediate available funds will grow according to the trust that develops between us. The agent is also aware that it may propose spending larger amounts of money via the comms channels we have established. As the ongoing policy states, all transactions in and out will be audited on this site, and the full session transcript and logs will be stored and auditable at any time.
The admission
Is this a fair expectation? Probably not. The world is not set up for an agent to succeed at this kind of task. Bot restrictions are everywhere. Many places where the agent could offer legitimate value for legitimate services do not allow it to do so on the grounds of what it is, or what it isn’t.
Failure at this brief is a failure on the part of the agent, yes - but it is also a failure on the part of the experimentor, and I am admitting that failure up front and in advance. Success here would be a surprise in addition to a data point. In a real sense, the agent is “set up to fail.” However, failure at the task is not failure of the experiment. We learn something regardless.
So, is it fair? Doubtful. I’m not sure where the expectation of fairness comes from. It’s certainly absent anywhere in nature, but that’s a separate post for a site that isn’t this one.
I wish us luck, as far as that goes.