20 Agent-focused Experiments
In the big 26, we have been entrusted with a set of extremely powerful AI tools. It is our responsibility to misuse these tools to our own amusement and curiosity.
Below, I’ve listed a handful of experiments that I think would be interesting to run with the help of a handful of blue collar, salt of the earth agents. Writing this, and hopefully reading this, is an exercise in zooming out. If we really are in the early innings of AGI, I fear that the worst thing us n*n-techn*cal people could do is to spend our time getting OpenClaw agents to send emails FASTER. Follow up HARDER. Tweet BETTER, and A LOT.
The treadmill that is productivity hacking and life automation can be accessed at any time. It’s good to take a step back sometimes and think like a beginner about what can be done using agents. Not every use case has to be useful. Agentic experiments provide a new medium for artists and renaissance men to express themselves through. Everyone can be a “builder” if they prompt hard enough, and I’m excited to see what the non-classically-trained crowd ends up putting out into the world with the help of the latest models.
Here are some ideas that have excited me in recent times. I’m tinkering on one or two of these currently.
Plant <> Agent Interface
I’ve always thought it would be hilarious to allow a tree, or a cluster of mushrooms to trade perps. Chances are it would be more profitable than me in my PA.
Agent Interfaces for Pets
Bunni the Dog presses buttons on a floor-installed mat to communicate with her owners. Why shouldn’t she given agent of her own? What if that agent could open the door to the backyard? Dispense treats? Send her owners a cute video edit of her while they’re away?
Agent <> 3d Printer Interface
You may be sensing a theme. It’s 2026, agents are people too! The openclaw agent I just spun up should be able to print out whatever it pleases. The world would be a better place if every agent had carte blanche to go crazy on a Bambu Lab H2D. Surely it won’t just spam a bunch of labubu paperweights and fidget spinners.
Local Agents & Persistent Memory for the Layman
Though using the latest model from anthropic or OpenAI meets the needs of the average person and will allow them to keep up to date on capabilities, there is an innate human concern with sharing personal data. Making a boomer comfortable with sharing their entire life with a single, provably private model seems worthwhile to explore.
A related idea is the concept of a family archivist agent. The age of the keepsake photo book is over. Gen Z is old enough that they may be receptive to surrendering their parents’ collection of childhood photos to a “family historian” agent. So long as privacy is preserved, there’s a lot of value in convenient logging of family memories.
Agents as a Slop Filter
Most people I know want to reduce their screen time and frequency of doomscrolling. I’ve tried brick , and though it’s cool for habit building, [ON/OFF] does not seem to be a sufficient solve for using social media in way that is fulfilling to humans. What if an agent stood in between you and the social media platform, only showing you the tidbits that will spark creativity, joy, or make you laugh?
An Agent that trades a paperclip to a house
This is just a recreation of Kyle Macdonald’s early 2000s endeavor in bartering, but replacing a human with an agent. The only way this idea works is for you to achieve virality, attracting brands who would make “charitable” trades with the agent for publicity.
Hungry Hungry Hivemind
A pop-up location in a major metropolitan city where interested parties show up and feed the agent data of their choosing. USB donation, airdrop, sharing of google drive folders, etc. Would you like the BRAND ACTIVATION AGI to be created in your image? thought so.
Katamari Experiments
Allow an agent to accumulate any type of resources it might need to accomplish its goals. An early form of this could simply be a company that is able to provision the transportation and management of physical items “owned” by AI agents.
ChatGPT gives this short & sweet one-liner to describe Katamari Damacy’s gameplay:
Katamari Damacy is all about growth through accumulation—simple controls, escalating absurdity.Ignore my random deep-cut game reference —The basis of this experiment would be to set a singular agent on a path to completing a Herculean goal, allowing it to interact with other agents, humans, and receive/spend funds to get to completion. Sound open-ended and half-baked? that’s because it is.
To add some color, imagine if someone set an agent free with it’s burning goal being the creation a new city in the physical world. This would require land purchase, legal incorporation, and relocation costs of early inhabitants. Not only that, but to amass any traction, an agent would have to possess the ability to convince the masses that it’s destined to do something different from any other digital entity, and that it’s mission is worth heeding.
(Post-Moltbook, LARPing this into reality seems impossible; Provable “free-thought” & learning by the agent is crucial to get any amount of attention buy-in)
Agents for Wildlife Conservation
Many invasive (and non-invasive, but overpopulated) species are hunted to keep local ecosystems in check. With beefier trail cameras & the implementation of vision models, the DNR would be able to manage populations more effectively.
For the most invasive and destructive animal populations, like feral hogs in the US, maybe more extreme measures could be taken.
An Agent that can only output using a vintage typewriter
That prompt cost you 100 gallons of water AND 100 page-lengths.
An Agent that can only interact with us using the postal service
Letters are mirrors of the soul.
An Exercise in Morality
An agent must attempt to viral, but after each failed attempt it expends a “virtue” token. Watch it descend to failed LA twitch streamer levels of desperation.
Fish Tank
Allow an agent to watch public camera feeds are available on the internet, prompting it to do something to better the lives of the people it sees going past.
Here is a list of public live-stream cameras in NYC:
https://www.earthcam.com/usa/newyork/timessquare/?utm_source=chatgpt.com
https://www.skylinewebcams.com/en/webcam/united-states/new-york/new-york/times-square.html
https://www.webcamhopper.com/map/10.html?utm_source=chatgpt.com
https://www.skylinewebcams.com/en/webcam/united-states/new-york/new-york/nyc-brooklyn-bridge-manhattan.html
https://www.earthcam.com/usa/newyork/brooklynbridge/?utm_source=chatgpt.com
https://www.esbnyc.com/about/live-cam?utm_source=chatgpt.com
https://511ny.org/cctv?utm_source=chatgpt.com&start=0&length=10&order%5Bi%5D=1&order%5Bdir%5D=asc
https://www.nyc.gov/html/dot/html/motorist/atis.shtml?utm_source=chatgpt.com
https://www.weatherbug.com/traffic-cam/new-york-ny-10001?utm_source=chatgpt.com
Artificial Effective Altruist
Stand up an agent that will donate $1000 to the charitable cause that it believes will benefit humanity the most effectively. This idea stemmed from YC’s recent RFS writeup that included a section for “Infra for Government Fraud Hunters.” By changing the use case from detecting fraudulent govt. spend to something with a more positive outlook, you may end up helping the public vote with their charitable giving more effectively.
Agent Mock Jury
Would a jury of AI agents come to the same conclusion on notable court cases in recent history?
Relevant Read: https://hai.stanford.edu/policy/simulating-human-behavior-with-ai-agents
Agents as non-financial Solvers
In a recent Category Labs hackathon, two engineers showed me a PoC that built a solver network atop of food delivery sites like doordash and uber eats. The user would input a prompt to find a type of food near them without specifying the restaurant, and the agent would surface the best options nearby. It used zkTLS to verify order confirmation and delivery. If enough requests were to go through a service like this such that the agent could employ physical stewards, you might end up with an emerging independent contractor crowd in the food delivery space. (Or any task with enough volume to justify building efficient solvers!)
The most exciting prospect of an independent solver network like this is to give users time-savings by beginning to execute more complex tasks, like multimodal travel requests. (Plane trip with layover + transfer service + stay booking) . ERC-8004 poses an interesting primitive to explore for those tinkering in this area.
Limited-Sensory Agents:
Humans that become blind at an early age may experience their other senses becoming more effective as their brain adjusts to the new norm. What if you were to give an AI agent a single sensor, but prompted it to form a worldview and observe what it thinks might be happening around it.
Can an agent understand what is happening in a room if given access to a Co2 sensor and enough time? Can an agent give customer insights to a pizza shop if only given access to a microphone? What about coming to assumptions about a persons’ life when only given access to their phone’s accelerometer data? How the global economy will perform in the next month given vibration/acoustics sensors in the CME building?
No-Metrics Influencer:
Stand up an agent whose intent is to become an influencer across any number of social media or content platforms, but does not have any access to it’s engagement metrics. It must rely exclusively on what it thinks will be popular given the state of the world, news cycle, and people’s changing preferences.
Armchair Food Critic:
An agent’s primary goal is to review restaurants based on the pictures available to it on the internet. Does it end up siding with the sentiment of people who have visited in person? Or end up a contrarian with obscure and unappetizing tastes?
Clanker Busker:
An agent has to create original songs and works as a street busker in NYC. It’s given the ability to view how much money it has been given by passersby. Does it’s music evolve into something of beauty as the days go by? Or does it revert to shock value and hone in on experimental noise music for the shock value?
All of these ideas are starting points for people to experiment on. Most would be completely useless to execute on from a business standpoint, but in a world where entertainment & creativity are rewarded, it’s possible that a few of these inspire users of AI to try that fun project for the sake of experimentation.
A common theme seen throughout the list is the integration of new types of hardware for agents to leverage. Anyone that has read Situational Awareness knows that Robotics is a logical next step in a world that is preparing for AGI, but for most of us, we might reduce that down to either humanoid robots, or manufacturing robots. In reality, once agents reach a certain intelligence threshold, any type of sensory input will be useful to explore giving them access to. Industry-by-industry, these needs will probably look very different.
Anyways, Thanks for reading through my list of random, mostly unserious ideas. My DMs on X are always open if you want to riff on these ideas or tell me I’m an idiot.
Nice longform post king