With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
Agent Builder (Bittensor Subnet 80) is a decentralized platform that allows users to create and customize their own AI agents by leveraging a network of miner-contributed AI models. In simple terms, it takes the best-performing AI agents contributed by miners on the Bittensor network and intelligently combines them to solve user requests. Unlike a basic chatbot, an agent built on this subnet can handle multi-step tasks, plan complex actions, and even utilize tools or web services as needed – all while learning from feedback to improve over time.
In practice, when a user sends a query or task to Agent Builder, an orchestrator component routes that request to one or more specialized miner agents best suited for the task. These miner agents (run by independent miners on the network) process the input using large language models (and potentially other AI tools) to produce responses. The orchestrator then ranks, refines, and aggregates the responses from multiple agents to construct the most accurate and helpful answer for the user. This means the final output can benefit from the combined expertise of several AI agents. The system is designed to support multi-turn interactions: an agent can perform a series of steps or a dialogue to complete a complex task, rather than just answering a single question. Agents can also store user preferences (e.g. remembering context or past instructions) via a unique session or content ID, enabling personalization and continuous improvement for repeat users.
Agent Builder (Bittensor Subnet 80) is a decentralized platform that allows users to create and customize their own AI agents by leveraging a network of miner-contributed AI models. In simple terms, it takes the best-performing AI agents contributed by miners on the Bittensor network and intelligently combines them to solve user requests. Unlike a basic chatbot, an agent built on this subnet can handle multi-step tasks, plan complex actions, and even utilize tools or web services as needed – all while learning from feedback to improve over time.
In practice, when a user sends a query or task to Agent Builder, an orchestrator component routes that request to one or more specialized miner agents best suited for the task. These miner agents (run by independent miners on the network) process the input using large language models (and potentially other AI tools) to produce responses. The orchestrator then ranks, refines, and aggregates the responses from multiple agents to construct the most accurate and helpful answer for the user. This means the final output can benefit from the combined expertise of several AI agents. The system is designed to support multi-turn interactions: an agent can perform a series of steps or a dialogue to complete a complex task, rather than just answering a single question. Agents can also store user preferences (e.g. remembering context or past instructions) via a unique session or content ID, enabling personalization and continuous improvement for repeat users.
Key capabilities of Agent Builder include:
Dynamic Agent Composition: It combines top-performing miner-contributed AI agents so that complex questions can be answered by the most capable models (or multiple models collaborating). This gives users a way to build an AI assistant that is stronger than any single model alone.
Tool Use and Multi-Step Reasoning: Miner agents aren’t limited to single-step Q&A – they can plan multi-step workflows. For example, an agent could break a problem into sub-tasks, call external tools or APIs (like web search, calculators, etc.), and then aggregate the results. The orchestrator coordinates this action–observation loop where the agent takes an action, observes the result, and decides the next step, enabling sophisticated autonomous behavior.
Learning from Feedback: The subnet has built-in feedback mechanisms. After responding to a query, agents receive feedback signals about their performance. This includes automatic scoring on metrics like answer quality, response speed (latency), reliability, and user satisfaction. There’s also a channel for human feedback – if a user provides a thumbs-up or correction, the agent can incorporate that. Over time, this feedback loop helps agents adapt and improve their answers, essentially self-improving with experience.
Decentralized and Incentivized: As a Bittensor subnet, Agent Builder operates without a central server. Independent miners contribute their AI models (agents) and compete to provide the best results. High-performing agents are rewarded with $TAO token emissions. This incentive model ensures continuous improvement, as miners have motivation to fine-tune and upgrade their agents to earn more rewards. The validators on the subnet evaluate miners’ outputs and help determine these rewards, keeping the ecosystem fair and performance-driven.
Agent Builder turns the Bittensor network into an “AI agent factory”. It gives users the power to deploy AI agents that can think and act in sophisticated ways, by harnessing a swarm of specialist models contributed by the community. This opens the door to AI assistants that can handle complex tasks autonomously – from researching information online, to executing multi-step workflows – all customized to the user’s needs and continuously learning from each interaction. It is essentially a decentralized application layer built on Bittensor where the “product” is an AI agent creation toolkit and service. Technically, the build consists of several components working together:
Miner Agents (AI Modules): These are the AI models provided by miners, each functioning as an autonomous agent. Miners implement their agents following a specified interface (with standard API endpoints) so they can communicate with the network. Each agent can receive tasks (/complete requests), generate responses, and handle feedback for continuous learning. The miner agents can be thought of as modular “skills” or sub-agents – one miner’s agent might be particularly good at coding tasks, another at answering medical questions, another at web browsing, etc. They run on the miners’ own hardware (GPUs/TPUs) and can be fine-tuned models, potentially lightweight (e.g. 7B–13B parameters) but optimized for tool use and multi-turn reasoning.
Orchestrator & Validator Nodes: At the heart of Agent Builder is an Orchestrator (this logic is usually implemented by the subnet’s validators). The orchestrator’s job is to take a user’s query and match it to the best agents, manage the multi-agent interaction, and synthesize the final result. It acts like a conductor: sending the query to one or multiple miner agents, waiting for their outputs, possibly asking some agents to refine their answers, and then choosing the highest-quality response or combining aspects of several responses. If the task requires multiple steps or tools, the orchestrator will loop through an action-plan: it may call an agent to perform step 1, pass the result to another agent for step 2 (/refine), and so forth, until the task is complete. Throughout this process, validators/orchestrator measure each agent’s performance on metrics like accuracy and speed.
Standardized Agent API: The product includes a well-defined API interface that every miner’s agent must support. There are four main API endpoints each agent implements:
These APIs ensure a common protocol so that many different agents (potentially written by different people in different languages) can all plug into the Agent Builder orchestrator seamlessly. The team has even provided a sample miner repository and a simple web UI (using Gradio) to help new miners get started quickly – with just a few commands, one can deploy a basic agent that connects to the subnet. This dramatically lowers the barrier to entry for builders, addressing the historically complex setup of Bittensor mining by offering a more plug-and-play template.
User Interface / Integration Layer: Although primarily a backend network service now, the vision for Agent Builder includes an accessible user interface or integration into applications. End users (or developers building on top) will interact with the orchestrator via a client (possibly a web app or API). For instance, a developer could integrate Agent Builder into a chat application or a decentralized app, allowing their users to spin up custom AI agents on demand. The platform might provide a no-code or low-code interface in the future, where a user can specify what they want their agent to do (choosing from available agent “skills” in the network) and then deploy an agent easily.
Behind the scenes, the Bittensor blockchain (Subtensor) supports all this by handling registration, reputation, and tokenomics. Miners register their hotkeys to join Subnet 80, paying a TAO stake to get in. The orchestrator (validators) then queries and scores them over time. TAO emissions are automatically distributed based on performance metrics – agents that consistently provide high-quality answers, fast responses, and reliable uptime get a larger share of the rewards. Poor performers may get pruned out or earn less, ensuring the “pool” of available agents remains competitive and effective. This competitive mining aspect is essentially an AI agent contest running continuously, which drives the overall product to improve.
In summary, the Agent Builder build is a combination of: a network of AI miner agents plus a coordinating orchestrator/validator system, all packaged into a cohesive platform where the end “product” is an AI agent that users can tailor to their needs. It’s like an app store of AI capabilities – but instead of downloading apps, you are invoking a network of AI agents and composing them into a custom super-agent. The platform handles all the heavy lifting of finding the right agents, merging their knowledge, and learning from feedback, delivering to the user a powerful AI assistant as the final product.
Key capabilities of Agent Builder include:
Dynamic Agent Composition: It combines top-performing miner-contributed AI agents so that complex questions can be answered by the most capable models (or multiple models collaborating). This gives users a way to build an AI assistant that is stronger than any single model alone.
Tool Use and Multi-Step Reasoning: Miner agents aren’t limited to single-step Q&A – they can plan multi-step workflows. For example, an agent could break a problem into sub-tasks, call external tools or APIs (like web search, calculators, etc.), and then aggregate the results. The orchestrator coordinates this action–observation loop where the agent takes an action, observes the result, and decides the next step, enabling sophisticated autonomous behavior.
Learning from Feedback: The subnet has built-in feedback mechanisms. After responding to a query, agents receive feedback signals about their performance. This includes automatic scoring on metrics like answer quality, response speed (latency), reliability, and user satisfaction. There’s also a channel for human feedback – if a user provides a thumbs-up or correction, the agent can incorporate that. Over time, this feedback loop helps agents adapt and improve their answers, essentially self-improving with experience.
Decentralized and Incentivized: As a Bittensor subnet, Agent Builder operates without a central server. Independent miners contribute their AI models (agents) and compete to provide the best results. High-performing agents are rewarded with $TAO token emissions. This incentive model ensures continuous improvement, as miners have motivation to fine-tune and upgrade their agents to earn more rewards. The validators on the subnet evaluate miners’ outputs and help determine these rewards, keeping the ecosystem fair and performance-driven.
Agent Builder turns the Bittensor network into an “AI agent factory”. It gives users the power to deploy AI agents that can think and act in sophisticated ways, by harnessing a swarm of specialist models contributed by the community. This opens the door to AI assistants that can handle complex tasks autonomously – from researching information online, to executing multi-step workflows – all customized to the user’s needs and continuously learning from each interaction. It is essentially a decentralized application layer built on Bittensor where the “product” is an AI agent creation toolkit and service. Technically, the build consists of several components working together:
Miner Agents (AI Modules): These are the AI models provided by miners, each functioning as an autonomous agent. Miners implement their agents following a specified interface (with standard API endpoints) so they can communicate with the network. Each agent can receive tasks (/complete requests), generate responses, and handle feedback for continuous learning. The miner agents can be thought of as modular “skills” or sub-agents – one miner’s agent might be particularly good at coding tasks, another at answering medical questions, another at web browsing, etc. They run on the miners’ own hardware (GPUs/TPUs) and can be fine-tuned models, potentially lightweight (e.g. 7B–13B parameters) but optimized for tool use and multi-turn reasoning.
Orchestrator & Validator Nodes: At the heart of Agent Builder is an Orchestrator (this logic is usually implemented by the subnet’s validators). The orchestrator’s job is to take a user’s query and match it to the best agents, manage the multi-agent interaction, and synthesize the final result. It acts like a conductor: sending the query to one or multiple miner agents, waiting for their outputs, possibly asking some agents to refine their answers, and then choosing the highest-quality response or combining aspects of several responses. If the task requires multiple steps or tools, the orchestrator will loop through an action-plan: it may call an agent to perform step 1, pass the result to another agent for step 2 (/refine), and so forth, until the task is complete. Throughout this process, validators/orchestrator measure each agent’s performance on metrics like accuracy and speed.
Standardized Agent API: The product includes a well-defined API interface that every miner’s agent must support. There are four main API endpoints each agent implements:
These APIs ensure a common protocol so that many different agents (potentially written by different people in different languages) can all plug into the Agent Builder orchestrator seamlessly. The team has even provided a sample miner repository and a simple web UI (using Gradio) to help new miners get started quickly – with just a few commands, one can deploy a basic agent that connects to the subnet. This dramatically lowers the barrier to entry for builders, addressing the historically complex setup of Bittensor mining by offering a more plug-and-play template.
User Interface / Integration Layer: Although primarily a backend network service now, the vision for Agent Builder includes an accessible user interface or integration into applications. End users (or developers building on top) will interact with the orchestrator via a client (possibly a web app or API). For instance, a developer could integrate Agent Builder into a chat application or a decentralized app, allowing their users to spin up custom AI agents on demand. The platform might provide a no-code or low-code interface in the future, where a user can specify what they want their agent to do (choosing from available agent “skills” in the network) and then deploy an agent easily.
Behind the scenes, the Bittensor blockchain (Subtensor) supports all this by handling registration, reputation, and tokenomics. Miners register their hotkeys to join Subnet 80, paying a TAO stake to get in. The orchestrator (validators) then queries and scores them over time. TAO emissions are automatically distributed based on performance metrics – agents that consistently provide high-quality answers, fast responses, and reliable uptime get a larger share of the rewards. Poor performers may get pruned out or earn less, ensuring the “pool” of available agents remains competitive and effective. This competitive mining aspect is essentially an AI agent contest running continuously, which drives the overall product to improve.
In summary, the Agent Builder build is a combination of: a network of AI miner agents plus a coordinating orchestrator/validator system, all packaged into a cohesive platform where the end “product” is an AI agent that users can tailor to their needs. It’s like an app store of AI capabilities – but instead of downloading apps, you are invoking a network of AI agents and composing them into a custom super-agent. The platform handles all the heavy lifting of finding the right agents, merging their knowledge, and learning from feedback, delivering to the user a powerful AI assistant as the final product.
One of the known core contributors is known by the handle “star145s.” This developer has been actively involved in building the Agent Builder infrastructure and supporting miners. For example, star145s created a public sample repository (agent-builder) that provides a template agent implementation and a friendly UI, making it easier for new miners to join the subnet. This indicates the team’s commitment to lowering barriers and engaging the community. On the project’s Discord (which has attracted over 130 members in its early days), star145s and presumably other team members actively interact with miners – fixing bugs, answering questions, and pushing updates. The communication from the team comes through frequent announcements and updates on X (Twitter) as well, often using a collective “we,” which suggests a small team collaborating rather than a single individual.
It’s worth noting that the TaoQuant team originally proposed Subnet 80 with a different concept (related to a decentralized fund for investment strategies), but they pivoted entirely to the Agent Builder concept in late 2025. This pivot highlights the team’s agile approach in pursuing what could have a bigger impact in the Bittensor ecosystem: enabling AI agents. Since the shift to Agent Builder, the team has been laser-focused on this AI agent platform, and the earlier concept is no longer being pursued. All official messaging now centers on AI agent development, with no overlap from the prior idea.
In terms of background, the team has not published biographies, but community speculation and the sophistication of Agent Builder’s design suggest they have strong machine learning backgrounds (possibly experience with LLMs and agent frameworks) and familiarity with decentralized systems. Some clues can be inferred from their work:
Implementing complex agent orchestration and support for things like the Berkeley Function Calling Leaderboard (BFCL) for tool use indicates deep knowledge of state-of-the-art AI research.
The ability to rapidly develop on Bittensor (which involves Rust/Substrate for chain aspects and Python for AI/model aspects) points to a technically skilled team that can straddle both AI and blockchain development.
Their proactive engagement with the Bittensor community and responsiveness to miner feedback reflect a professional approach, albeit under pseudonyms.
One of the known core contributors is known by the handle “star145s.” This developer has been actively involved in building the Agent Builder infrastructure and supporting miners. For example, star145s created a public sample repository (agent-builder) that provides a template agent implementation and a friendly UI, making it easier for new miners to join the subnet. This indicates the team’s commitment to lowering barriers and engaging the community. On the project’s Discord (which has attracted over 130 members in its early days), star145s and presumably other team members actively interact with miners – fixing bugs, answering questions, and pushing updates. The communication from the team comes through frequent announcements and updates on X (Twitter) as well, often using a collective “we,” which suggests a small team collaborating rather than a single individual.
It’s worth noting that the TaoQuant team originally proposed Subnet 80 with a different concept (related to a decentralized fund for investment strategies), but they pivoted entirely to the Agent Builder concept in late 2025. This pivot highlights the team’s agile approach in pursuing what could have a bigger impact in the Bittensor ecosystem: enabling AI agents. Since the shift to Agent Builder, the team has been laser-focused on this AI agent platform, and the earlier concept is no longer being pursued. All official messaging now centers on AI agent development, with no overlap from the prior idea.
In terms of background, the team has not published biographies, but community speculation and the sophistication of Agent Builder’s design suggest they have strong machine learning backgrounds (possibly experience with LLMs and agent frameworks) and familiarity with decentralized systems. Some clues can be inferred from their work:
Implementing complex agent orchestration and support for things like the Berkeley Function Calling Leaderboard (BFCL) for tool use indicates deep knowledge of state-of-the-art AI research.
The ability to rapidly develop on Bittensor (which involves Rust/Substrate for chain aspects and Python for AI/model aspects) points to a technically skilled team that can straddle both AI and blockchain development.
Their proactive engagement with the Bittensor community and responsiveness to miner feedback reflect a professional approach, albeit under pseudonyms.
The Agent Builder roadmap is centered on rapidly evolving the platform from its initial launch (v1) into a robust, user-friendly ecosystem for AI agents. Although the team hasn’t published a formal public roadmap document, they have communicated several milestones and future plans through updates. Here’s what can be expected moving forward:
Current Phase – V1 Mining Contest and Stability: (Late Oct 2025 – Present) Agent Builder launched its v1 mining phase at the end of October 2025. This phase kicked off with a “Miner Contest”, essentially an open invitation for miners to join Subnet 80 and start building agents. In the first two weeks, the subnet saw strong engagement (130+ Discord members, nearly 40 miners actively working on agents). The initial goals in this phase are:
Short Term – Ramp Up and Feature Completeness: (Late Q4 2025) After the initial contest period and once the subnet proves stable, Agent Builder is likely to increase the emissions gradually towards 100%. This means miners will earn more, attracting even more participants and incentivizing improvements. We anticipate:
Medium Term – User-Facing Platform & Integration: (Q1–Q2 2026) Once the agent network is robust, Agent Builder will likely turn toward the end-user experience. The vision is to empower not just miners and developers, but also non-technical users to spin up AI agents. Possible roadmap items in this stage:
Long Term – Mature Ecosystem & Governance: (Late 2026 and beyond) In the long run, Agent Builder aims to be a foundational layer for decentralized AI services. Some forward-looking aspects of the roadmap might be:
While many of these future steps are aspirational, the trajectory is clear: Agent Builder is quickly moving from an experimental subnet into a full-fledged Agent Platform. The team’s fast execution in the contest phase and continuous improvements suggest that new features will roll out frequently. We can expect public updates from the team at each milestone (they have been actively sharing progress on social media). By monitoring those, one can see the roadmap unfold in real-time. The excitement in the community is high – Agent Builder is viewed as a potential game-changer in the Bittensor ecosystem, and the coming months will be focused on turning that potential into reality.
The Agent Builder roadmap is centered on rapidly evolving the platform from its initial launch (v1) into a robust, user-friendly ecosystem for AI agents. Although the team hasn’t published a formal public roadmap document, they have communicated several milestones and future plans through updates. Here’s what can be expected moving forward:
Current Phase – V1 Mining Contest and Stability: (Late Oct 2025 – Present) Agent Builder launched its v1 mining phase at the end of October 2025. This phase kicked off with a “Miner Contest”, essentially an open invitation for miners to join Subnet 80 and start building agents. In the first two weeks, the subnet saw strong engagement (130+ Discord members, nearly 40 miners actively working on agents). The initial goals in this phase are:
Short Term – Ramp Up and Feature Completeness: (Late Q4 2025) After the initial contest period and once the subnet proves stable, Agent Builder is likely to increase the emissions gradually towards 100%. This means miners will earn more, attracting even more participants and incentivizing improvements. We anticipate:
Medium Term – User-Facing Platform & Integration: (Q1–Q2 2026) Once the agent network is robust, Agent Builder will likely turn toward the end-user experience. The vision is to empower not just miners and developers, but also non-technical users to spin up AI agents. Possible roadmap items in this stage:
Long Term – Mature Ecosystem & Governance: (Late 2026 and beyond) In the long run, Agent Builder aims to be a foundational layer for decentralized AI services. Some forward-looking aspects of the roadmap might be:
While many of these future steps are aspirational, the trajectory is clear: Agent Builder is quickly moving from an experimental subnet into a full-fledged Agent Platform. The team’s fast execution in the contest phase and continuous improvements suggest that new features will roll out frequently. We can expect public updates from the team at each milestone (they have been actively sharing progress on social media). By monitoring those, one can see the roadmap unfold in real-time. The excitement in the community is high – Agent Builder is viewed as a potential game-changer in the Bittensor ecosystem, and the coming months will be focused on turning that potential into reality.