Meet the Agents That Pay for Their Own Compute: Inside Aeon, MiroShark, and Agentic Commerce

What happens when software stops waiting for you to ask? Most AI agents today fall into two camps: chat wrappers that sit idle until prompted, and rigid bots that act without any judgment. Aaron Elijah Mars is building for the gap in between. His two open-source projects, Aeon and MiroShark, share a single conviction, that software should be able to run, observe, and ship on its own. \ In this conversation we get into autonomous agents that file their own pull requests, synthetic crowds that trade against a live market, and what agentic commerce actually looks like when the agents are also the builders. \ Ishan Pandey: Hi Aaron, it's a pleasure to welcome you to our "Behind the Startup" series. Please tell us about yourself and what inspired you to build Aeon and MiroShark? \ Aaron Elijah Mars: Hi Ishan! Thanks for having me. I’ve been thinking about the concept of aeon for the past year; I wanted my Claude to be more proactive, sending me ideas, good content to read or digest based on what I was interested in. But it was pre-OpenClaw & most of the agent design was very simple, so I’ve built a few prototypes & stopped there. Fast forward to 3 months ago, I realized it was a good time to work on this idea again & try to port Hermes to use a similar proactive approach. Quickly realized that it wasn’t possible with their stateful architecture, so I’ve designed aeon to maximize autonomy & picked GitHub as the main source of truth + execution. \ For MiroShark it was more accidental. I’ve previously built a prediction market called Hyperstitions, that focuses on motivating people to do a specific action / task instead of just ‘pricing the truth.' After running a few markets, I realized that: 1) it’s really hard to predict people's behavior. 2) but it’s also ultra-valuable, especially for AI to make decisions. \ MiroShark is a simulation company; think World Models but for decision-making. You can ask any question, have relevant personas come up on the simulation, add a new event on the simulation, see how everyone reacts, etc. \ Ishan Pandey: Your thesis is that software shouldn't wait for a human prompt to run. What broke for you that made "wait for a prompt" the thing you wanted to design against? \ Aaron Elijah Mars: People love to compare humans & AI. One of the main differences between both is that humans are data-intensive, we always grab data to fine-tune our internal models. If I know you - I’m aware of how you’ll behave in a particular situation & how I should speak to you. If I see you sad, I’ll also adapt. AI doesn’t do any of this. The second difference is that humans usually don’t wait for approval to do things. You close your eyes, the world keeps moving. AI doesn’t. \ Ishan Pandey: Aeon positions itself between chat wrappers you have to ask and rigid bots with no judgment. Where does that judgment actually live in a skill, in the SKILL.md, the model, the schedule, or the person composing them, and what did you have to trade away to make the full agent log public and auditable by default? \ Aaron Elijah Mars: Correct. We’ve built three things that are important : first an aeon has a SOUL.md. Usually based on you & how you behave, ingesting all your data & making a mental model of who you are / what you like & what you don’t. Which means it filters for your worldview automatically. Second is your STRATEGY.md , which builds a mental model of what you want to achieve & lets your aeon optimize for it. The last part is a byproduct of the design. Everything runs on GitHub, which means everything is auditable in real time. If your aeon builds a new skill, you can understand why. If it breaks, you can roll back. \ Ishan Pandey: You draw a hard line between MiroShark being a simulation rather than a survey, where agents see each other, change their minds, and trade against a working market. Can you break down how the three coupled surfaces work together, and why that produces a signal a single-agent "AI focus group" tool structurally cannot? \ Aaron Elijah Mars: The biggest issue with AI focus group tools is diversity. When you ask a question to different Anthropic models, they pretty much always have the same answer / worldview on it. When you run MiroShark, you feed it an article / corpus that enables it to create agents related to the article. Feed it an article on SpaceX, it will create an Elon Musk persona, a Cursor persona, etc. Each agent gets a deep personality, different models & most of all, ways to act on the simulation. When enough context is gathered, all agents enter simulation mode, which requires them to do actions (or not) on a simulated X, Reddit & Polymarket. They see each other’s actions & react to them. If Elon Musk decides to sell his Yes shares on Polymarket, it impacts the posts on X & Reddit. This is way more chaotic & situational than a ‘snapshot in time’ system that just asks AI opinion about something. It’s not just ‘what do you think about this wording’, it’s ‘here is what could happen if you post this’. \ After the simulation happens, we condense the interesting parts, cut the noise & give you what really matters. \ Ishan Pandey: MiroShark runs on x402 over Base with no API key or signup, and you've put work on Bankr, where agents fund their own inference through swap fees on a token they launch. How does "the agent pays for its own compute" change the economics and the design of what you build, and what does it unlock for agent-to-agent commerce that a traditional API-key model cannot? \ Aaron Elijah Mars: x402 is one of the most important design pieces of the architecture. If you still need a human operator to add new capabilities, you are not building the most autonomous agent framework. It truly enables the agent to self-evolve & add new capabilities. In terms of business model, it also changes everything, where each new agent skill can become a product. And this product can help pay for compute in a virtuous circle that transforms an agent into a full company. \ Ishan Pandey: A working prediction market with a real AMM sitting inside a simulation invites an obvious challenge: can a run be gamed, either by a client who wants a flattering result or by the agents themselves drifting into a bubble? You also merged a two-model external security audit same-day. How do you guard the integrity of the output, and what was the most uncomfortable finding? \ Aaron Elijah Mars: It’s the beauty of the world. Truth is usually way more reflexive than we think, bubbles shouldn’t exist yet they are part of our economy, etc. We try to simulate hard problems, that involve humans, chaos theory & sometimes greed. Prediction markets are a great way to aggregate the truth, but sometimes you don’t have the budget to create the market & have enough people to care about it. So we simulate it, paying $1 instead of millions. \ Ishan Pandey: What advice would you give founders building open-source autonomous agents, on balancing speed, trust, and a business model that doesn't lock the work away? \ Aaron Elijah Mars: You need to be your first user and be deeply interested in the use case it enables. Claude Code showcased to the whole world that the most valuable part of the stack is the harness. Especially for coding. Now how does a harness for other domains look? What’s the form factor? I guess that’s the most important question right now. \ Don’t forget to like and share the story! :::tip Vested Interest Disclosure: HackerNoon has reviewed the report for quality, but the claims herein belong to the author. #DYOR. ::: \n \