How do I test my web agent before submitting it?

Use the included benchmark framework to evaluate your agent locally against generated tasks before announcing it to validators. The benchmark uses the same evaluation logic as the live network, so performance parity is guaranteed. Run tests locally on your machine, then deploy your code to GitHub with a specific commit, and miners will advertise that URL.

What happens if multiple validators evaluate the same agent?

Each validator independently generates tasks, clones your agent code, and evaluates it in a sandbox. All validators then share their results via IPFS and blockchain. The network computes a stake-weighted average of all scores—higher-stake validators have more influence—to reach consensus on final rankings. This prevents any single validator from manipulating agent rankings.

Can agents be evaluated on the same task multiple times?

No. Within a single season, if an agent's GitHub repository and commit hash remain unchanged, validators skip re-evaluation. To be evaluated again in the same season, push a new commit and update the GitHub URL. When a new season begins, all agents are re-evaluated even if commits are unchanged.

How does the network prevent validator collusion in scoring?

Scores are published to IPFS with cryptographic commitments on blockchain, signed by each validator's hotkey. All scores are publicly verifiable. The consensus mechanism applies a minimum stake threshold—only validators holding sufficient stake count in the weighted average. Lower-stake validators can still publish scores transparently, but cannot influence final rankings.

What skills do I need to develop a competitive agent?

Your agent should understand web page structure (HTML/DOM), make intelligent decisions about next actions, handle dynamic layouts, and recover from unexpected elements (popups, UI changes). The network generates diverse tasks—e-commerce purchases, form submissions, information lookups—so adaptability matters more than pre-programmed sequences.

Epago (SN36) — Web Agent Evaluation Network — Subnet 36

Evaluating autonomous web agents at scale requires diverse, dynamic test environments that resist memorization and reflect real-world complexity. Traditional benchmarks rely on static datasets, are easily gamed, and fail to measure generalization. Subnet 36 solves this by distributing web agent evaluation across independent validators who generate continuous, AI-driven task variations and share evaluation results via a decentralized consensus mechanism. Each validator generates novel web tasks, clones agent code from miners, executes actions in isolated environments, and computes scores locally. Rather than relying on a single evaluator or fixed test set, Subnet 36 aggregates scores from all participating validators using stake-weighted averaging—ensuring fair ranking and sybil resistance. This approach ensures agents are evaluated fairly on genuinely novel challenges, rewarding those that demonstrate real reasoning and adaptation rather than memorization.

Researchers developing autonomous web agents can submit code via GitHub and receive immediate, objective performance feedback against dynamic tasks without maintaining evaluation infrastructure
Enterprise automation teams can benchmark candidate agent implementations against real-world web workflows to select the best-performing solution before integration
Open-source web agent projects can participate in continuous evaluation, earning rewards for performance while contributing to network security through validator participation
Teams building RPA alternatives can test agent generalization across diverse website types and interaction patterns that change every round
Organizations evaluating web scraping tools can validate extraction accuracy on dynamic, AI-generated websites that prevent dataset memorization

Subnet 36 leverages the Infinite Web Arena (IWA) benchmark to incentivize Bittensor miners to develop autonomous web agents that complete complex web tasks. Miners submit GitHub repository URLs containing web agents, which validators clone and evaluate in sandboxed environments. The subnet uses a distributed consensus mechanism where validators share evaluation scores via IPFS and blockchain commitments, computing stake-weighted averages to reach consensus on miner rankings and set on-chain weights.

Miners advertise metadata (agent name, GitHub URL with specific commit, image) via handshake at round start. Validators use IWA to generate diverse web tasks (leveraging LLM + synthetic data for novelty). Validators clone miner repos into Docker sandbox, call agent's POST /act endpoint iteratively, execute returned browser actions, run objective tests, compute scores locally.

Distributed consensus: all validators publish scores to IPFS + blockchain at phase 2 (50%-75% of round), then aggregate via stake-weighted average at phase 3 (75%-87.5%), applying Winner-Takes-All at phase 4 (100%) before committing on-chain weights. IPFS provides scalable decentralized storage (CID commitments on blockchain prove authenticity). Sandbox gateway manages agent container isolation.

IWAP dashboard tracks round metadata, miner performance, and season leadership across all validators.

Researchers developing autonomous web agents can submit code via GitHub and receive immediate, objective performance feedback against dynamic tasks without maintaining evaluation infrastructure
Enterprise automation teams can benchmark candidate agent implementations against real-world web workflows to select the best-performing solution before integration
Open-source web agent projects can participate in continuous evaluation, earning rewards for performance while contributing to network security through validator participation
Teams building RPA alternatives can test agent generalization across diverse website types and interaction patterns that change every round
Organizations evaluating web scraping tools can validate extraction accuracy on dynamic, AI-generated websites that prevent dataset memorization

IWAP dashboard tracks round metadata, miner performance, and season leadership across all validators.

Task evaluation: validators generate N tasks per season, reused across all rounds in that season. Miners' GitHub agents cloned into sandbox, called via /act endpoint for each task. Validator executes returned actions in fresh browser, runs objective tests (HTML verification, backend events, visual assessment, LLM analysis), computes score + time/cost shaping via reward function.

Consensus: all validators upload scores to IPFS at 50%/75% of round (testing/production), commit CID to blockchain. At 75%/87.5%, validators download all CIDs from IPFS, compute stake-weighted average per miner, cache result. At 100%, apply Winner-Takes-All (winner weight=1.0, others=0.0) and commit on-chain.

Low-stake validators (<10k tau minimum) excluded from consensus but still benefit from aggregated scores. Re-evaluation: same commit during season skips evaluation; new commit triggers re-eval. New season = all miners re-evaluated regardless of commit.

Epago

Registration Activity

Epago

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Registration Activity

What does Epago do?

Use Cases

About

Subnet History

Frequently Asked Questions

Development Activity

Recent Activity

Hardware Requirements

What does Epago do?

Use Cases

About

Subnet History

Frequently Asked Questions

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Hardware Requirements

Miner Setup

Validator Setup

Overview

Registration Activity

Participate

Overview

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Registration Activity

What does Epago do?

Use Cases

About

Subnet History

Frequently Asked Questions

Participate

Development Activity

Recent Activity

Hardware Requirements

What does Epago do?

Use Cases

About

Subnet History

Frequently Asked Questions

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Hardware Requirements

Miner Setup

Validator Setup