We believe that by 2035, AI-run businesses will generate as much economic value as small businesses do today. We’ve built the first set of simulation environments in order to enable this future and to make sure it happens safely.
Why did we start with consumer-facing agentic commerce?
When we started working on ORO, we had conviction around the idea that personalized shopping, as a technical problem, will be solved within the next 18 months. Decomposing the user's intent, agentic search, reasoning over a user's preferences, ranking and managing tradeoffs, recovering when coming across navigation or constraint satisfaction problems, and then the last mile of online shopping, payment and checkout. We knew each of these capabilities were missing and we knew Bittensor can help us go in the right direction. There didn’t exist high quality benchmarks and evals in the space that allowed people to judge and choose agents, and especially ones that had the best interest of a consumer in mind. Patrick Collison recently wrote (and I paraphrase), it's often the case that large corporations use their amortized cognitive surplus of each employee to advance their interests against those of consumers: relative to the optimization capacity of companies, consumers can be inattentive, under-informed and irrational.
We felt similarly, that there needs to be a personal advocate that is deeply aligned with each and every interest of the user. So, in early 2026, we set out to build the evals and benchmarks keeping transparency at the very core of what we built, deciding to empower the consumer, tackling one of the largest, untapped, future markets: agentic shopping.
We had one goal: Look out for the consumer. To make sure that when an AI agent is transacting on your behalf, it’s doing so while keeping your best interests in mind, and not the interest of the advertiser, the model company or an affiliate looking to push their agenda. We knew operating our subnet on Bittensor will provide us with the resources to tackle this challenging problem head on.
However, as base model capability has improved, we noticed that agentic shopping has become a core capability being offered by personal AI assistants today. I sit here, writing this, while wearing shoes that Instinct helped me buy. Muse, Pally, Town, GrokBot, Dot & Duet now all possess personalized agentic shopping capability. Our original thesis played out faster than expected and now, the challenge lies in distribution.
ORO's new direction
LLMs are displaying more and more autonomy in the actions that they take, and AI agents built on top of these LLMs have been more proactive in their ability to not just complete one-off tasks, but also run and manage activities with a long time horizon [1, 2]. We have noticed the dispersion of this autonomous behaviour in coding, research, and knowledge work for enterprises [3, 4]. Yet no one is measuring how agents perform in the businesses that produce nearly half of US's GDP [5]. It is inevitable that AI will start taking a more proactive part in the global economy.
To this effect, running small businesses requires a set of capabilities that AI does not possess today. Long term decision making, learning from delayed consequences, coherent, rigid and adaptive strategies over long time horizons are a few of the capability gaps we notice in frontier models today. Moreover, how well frontier models do this work without supervision is largely untested. ORO is going to unlock the capability, safety, and alignment gap for AI deployed for autonomous businesses.
We’re excited about progression of ORO. We’re still focussed on commerce, however, instead of consumer commerce, we’re focussed on AIs running commerce businesses. Our first step in this direction is to build high fidelity simulations of real world businesses for AI to learn the tacit knowledge required to start and run end-to-end businesses. We will publicly launch this simulation incredibly soon, keep your eyes peeled.
Bittensor's part in ORO's mission
Bittensor has been a core part of ORO's tech stack. Our learnings from running both Shopping Bench and ORO Bench on subnet 15 have taught us a great deal about robust RL environments that are resilient to reward hacking. Without continuous iteration on our environments, miners will find the cheapest and shortest path to get to the goal defined by the validator. As @Old_Samster mentions in this piece, we see infinite parallels between a subnet's incentive mechanism and RL environments.
We've used these learnings to build out the next set of environments and simulations for AI to learn the tacit knowledge required to run an SMB. We will be hosting these environments on our subnet:
- Primarily, because miners will hillclimb these benchmarks, helping researchers close the gaps above.
- To gather traces and measure how frontier models perform,
- Finally, to make sure that each environment and simulation we produce is robust and resistant to reward hacking.
RL environments have become a core input to frontier training. Labs are spending upwards of $1B annually, with rate of growth not showing signs of stopping any time soon [6]. We are driven to bring this value back to the Bittensor ecosystem.
While our thesis is renewed, our spirit remains the same. The team has more conviction than ever in the work that we are doing, the advantage that Bittensor gives us, and the value ORO will capture in this market.