sundae_bar
SN121AI agents that handle real business workflows end-to-end, from email to execution
A London-listed company using Bittensor to build and commercially deploy a generalist AI agent through competitive mining. The subnet trains the agent, the enterprise platform sells it, and revenue flows back to the network.
// One agent. Continuously improving. Commercially deployed.
sundae_bar (SN121) is a Bittensor subnet designed to build and continuously improve a single generalist AI agent through open competition. Developers submit agents, validators benchmark them, and the best-performing agent gets deployed commercially through the sundae_bar enterprise platform.
The simple version: It's like a continuous public competition where developers race to build the best general-purpose AI assistant, and the winner gets deployed to paying business customers.
Centralized equivalent: OpenAI's GPT-4 for enterprises, or Microsoft Copilot. The difference is that sundae_bar's agent is built openly by competing developers rather than a closed internal team.
How it works:
- Miners (Developers) submit fully open-source generalist agents, evaluated against Generalist Challenges defined by the sundae_bar team
- Validators run the Agent Eval Test Suite (AETS) across multiple seeds and configurations, aggregating scores into a consensus performance ranking
- The problem it solves: Building a reliable generalist AI agent requires enormous compute, diverse evaluation, and continuous iteration. No single team can do this as efficiently as an open competitive market.
- The opportunity: Enterprise workflow automation is a large and growing market. sundae_bar is betting that competitive, open development produces a better agent faster than closed models.
- The Bittensor advantage: Every improvement is immediately benchmarked, rewarded, and deployed. All agent code is open-source. The competitive pressure is continuous rather than periodic.
- Traction signals: sundae_bar is a publicly listed company on the London AIM exchange (ticker: SBAR). The subnet README describes a revenue-backed emissions model where commercial revenue from renting the agent triggers buybacks, capping miner rewards at 70% of realized revenue.
Category: Reinforcement Learning | Centralized Competitor: OpenAI GPT-4, Microsoft Copilot, Anthropic Claude for Enterprise
sundae_bar takes a direct commercial approach. The subnet is not building infrastructure or a niche tool; it is building a product for businesses that need AI-powered workflow automation, and using Bittensor's competitive mining structure to do it faster and more transparently than a closed R&D team could.
Mechanism:
According to the subnet's README, sundae_bar defines Generalist Challenges: structured multi-domain datasets with evaluation rubrics and task hierarchies. Miners submit open-source agents built on compatible frameworks (initially Letta, with LangChain, AutoGen, and CrewAI planned). Validators run the Agent Eval Test Suite across multiple seeds and configurations. Results are aggregated into a consensus performance score, and emissions go to the top-performing agent in a winner-takes-all structure within each evaluation window.
The economic model is designed to be self-sustaining. Before commercial revenue, the subnet runs on controlled low emissions with unused emissions burned. Once business customers pay to rent the agent through the enterprise platform, 70% flows to miner rewards and the remainder strengthens the liquidity pool. This creates a direct link between the quality of the on-chain agent and the subnet's economic health.
One significant flag: the subnet currently has 0 active miners, 0% emission share, and net outflow, meaning it is not receiving network emissions under the current Taoflow model. The README notes this is expected in the pre-revenue phase, with a controlled emissions bootstrap, but it is worth noting for anyone evaluating staking.
- Zero emissions, zero miners: Under Taoflow, the subnet currently receives no network emissions and has no active miners. It is entirely in pre-revenue bootstrap mode.
- Execution: The commercial revenue model depends on businesses actually paying for the agent. That has not happened yet based on available data.
- Winner-takes-all risk: The evaluation structure rewards the single best agent. This can create fragility if the top miner exits or the agent quality plateaus.
- Concentration: Gini of 0.78 and HHI of 0.14 indicate a fairly concentrated holder base.
- Development: The GitHub repo shows 12 commits from 2 contributors, with the last commit in December 2025. Low development activity for a subnet with ambitious commercial goals.