36 changed hands. The slot that used to run Autoppia's web agents now hosts Eirel, an open competition to build the best AI assistant, where the chain grades each entry on whether it actually got the answer right.
//What is Eirel
Eirel is a Bittensor subnet where independent developers compete to build AI assistants. Each developer packages their assistant as a self-contained program, submits it, and the network runs it against a battery of tasks and scores how well it answers. Rewards follow the scores.
The simple version: It is a bake-off for chatbots. Everyone enters their own assistant, the same questions go to all of them, and the entries that answer best, and use their tools correctly, earn the rewards.
Centralized equivalent: Think ChatGPT or Perplexity, but instead of one company building the assistant behind closed doors, anyone can submit one and an open scoring system decides which is best.
How it works:
submit a Docker-packaged conversational agent that can chat over multiple turns, search the web, fetch URLs, run Python in a sandbox, and pull from documents.
Validators run each miner's agent on evaluation tasks and score the answers, gating on whether the response is grounded and correct, whether it stayed safe, and whether the agent's tool use matches a server-side log of what it actually called.
//Why This Matters
The problem it solves: Most assistant benchmarks are run by the same people who build the model, and "it sounds right" is easy to game. Eirel separates building from grading. The subnet operator runs the scoring plumbing, and validators judge independently against tasks the miner did not write.
Keep exploring
Other research from the same neighborhood of the network.
The opportunity: A continuously run, open contest for assistant quality, where the scoring leans on grounded correctness and verified tool use rather than a static leaderboard that is simple to overfit.
The Bittensor advantage: The work (building a good assistant) and the judging (scoring it) sit with different parties whose incentives are set on-chain, rather than inside one company that benefits from its own model winning.
Traction signals: Eirel is early. The SN36 slot was re-registered under this team in April 2026, the code stack has been pushed steadily through late May, and the first competition family is the one now coming online. There is little public social footprint yet.
//Full Analysis
Category: Other (Conversational AI Agents) | Centralized Competitor: ChatGPT, Perplexity, Gemini
Subnet 36 is a clean example of how a Bittensor slot can change occupants. It previously ran Autoppia, a network for autonomous web agents. On-chain, the slot now resolves to a different owner and a different project: Eirel, which describes itself as "the execution layer for multimodal AI workflows." When a subnet is deregistered, its netuid can be claimed by a new registration, and that is what happened here. This article covers the current occupant, verified against Eirel's own repository.
Mechanism:
Eirel runs a competition with a clear division of labor, taken from the project's `eirel-ai` repository. The subnet owner operates a control plane (the repo calls it `owner-api`) that handles submissions, builds and runs miner code, orchestrates evaluation, aggregates scores, and publishes weights to the chain. The owner does not judge. Each validator runs its own judging sidecar.
The lifecycle, per the repo:
A miner submits a Docker-packaged agent along with a small on-chain fee (0.1 TAO in the documented configuration), and the submission's signature is verified through Bittensor.
The control plane builds the image and places it on subnet-owned runtime, either Docker or Kubernetes.
At each run boundary, the current set of deployed agents is frozen and evaluation tasks are seeded across a matrix of capabilities and domains, so coverage stays broad.
Validators claim tasks, invoke each miner's agent through the operator's proxy, and judge locally. The documented scoring is a composite that gates on grounded correctness and instruction safety, and it cross-checks the agent's tool calls against a server-side ledger rather than trusting what the agent reports about itself.
Scores roll up per run into aggregate that drive weight publication. After a run closes, the submission archives that were scored become publicly downloadable from the leaderboard, so operators can audit and competitors can study what won.
The launch family is `general_chat`: a multi-turn conversational assistant offering both a quick mode and a slower reasoning mode, backed by owner-routed tools for web search, URL fetch, a Python sandbox for verifiable computation, and retrieval over per-run document sets. The repository lists additional families as roadmap items gated behind a launch flag, so the current scope is conversational assistants first, with more task types planned.
A design detail worth calling out: the tool-call attestation. Because validators score the agent's tool use from the operator's own log of what was called, not from traces the agent emits, a miner cannot claim it searched the web or ran code that it did not actually run. That is a sensible answer to a real problem in agent evaluation, where self-reported behavior is easy to fake.
On the numbers, Eirel is small and recent. Alpha trades around 0.00533 TAO with a near 3,850 TAO, and the pool holds roughly 1,290 TAO in root . Its smoothed share sits under one percent. is high, around 0.58, which is typical of a subnet whose price has not settled after a recent (re)launch. Net staking flows have been modestly negative over the past week across the data sources checked, and the alpha price is down roughly 35 percent over the last 30 days, spanning the relaunch period. Development is live: the `eirel-ai` control-plane repository, plus the related `eirel` and `eiretes` packages it depends on, were all pushed within the last few weeks, with the most recent control-plane commits landing a run of security fixes around the miner-isolation and token-scoping logic.
One caveat on sources. A third-party about page for this slot still describes the previous Autoppia project. That is stale data lagging the on-chain change, not a second version of Eirel. Where the third-party copy and the live repository disagreed, this article follows the repository.
//Risk Factors
These factors move fast; captured at publishing date
Execution risk: Eirel is new to this slot and unproven at scale. The slot was re-registered in April 2026, the control-plane repository is carried mainly by a single maintainer, and the first competition family is only now coming online. Working code is not the same as a working market of miners and validators.
Market risk: The alpha price is down roughly 35 percent over 30 days, and net staking flows have been slightly negative over the past week. Under the current flow-based emission model, sustained negative pull a subnet's emission share down, so the inflow trend is worth watching.
Competition: Conversational assistants are among the most crowded categories anywhere, against well-funded centralized products and other agent-focused Bittensor subnets. Eirel has to show its open, attested scoring produces assistants that are actually better, not just differently governed.
Concentration: A gini coefficient around 0.70 points to concentrated stake distribution. Large positions can move pool dynamics and weight outcomes more than a flatter distribution would.
