Harnyx
SN67Inactive or reserved subnet slot
A subnet that turns deep research into a competitive harness, with miners submitting Python agents that get sandboxed, scored against reference answers, and replaced only when a challenger is materially better.
// Better harnesses compound faster than better models.
Harnyx (SN67) is a Bittensor subnet for deep research. Miners write Python agents that answer research-style questions under a tight tool budget, validators run those agents in sandboxes against benchmark tasks, and the network keeps whichever agent is currently best as the "champion."
The simple version: It's like a continuous coding competition for research bots, where the winner stays on the throne until a challenger clearly does better.
Centralized equivalent: Think OpenAI Deep Research or Perplexity Pro, except the actual research workflow is built and improved by anyone willing to compete, and every run can be inspected.
How it works:
- Miners submit Python agent scripts that take a query and return a research answer, while staying inside a tool budget
- Validators execute those scripts in sandboxed containers against a stream of tasks and score the outputs against a stronger reference answer
- The problem it solves: Deep research is not one reasoning step. It is decomposition, retrieval, ranking, cross-checking, and synthesis. Closed product teams iterate on those harnesses behind a wall, and the wider field cannot tell what is actually working.
- The opportunity: If the harness around a model matters as much as the model itself, an open arena where harnesses are continuously stress-tested is a real lever on research quality, not a cosmetic one.
- The Bittensor advantage: Sandboxed execution, a public benchmark, and on-chain emission make the competition both legible and economically motivated. Champion replacement rules push for genuine improvement, not noise.
- Traction signals: The repo went live in March 2026 and shows daily commits, with the most recent push on May 12, 2026. Live benchmark history is published at dashboard.harnyx.ai/benchmark, and miner and validator guides are in the repo.
Category: Search and Information Retrieval | Centralized Competitor: OpenAI Deep Research, Perplexity Pro, Exa
Deep research as a product category took off in 2025, and most of the interesting work happens inside closed labs. Harnyx is one of the first attempts on Bittensor to make the research harness itself the asset, with the model layer treated as one ingredient among many. The thesis is that compounding on harness quality, decomposition, retrieval shape, ranking, cross-checking, beats compounding on raw model size for this task.
Mechanism:
A task is one research-style query plus a stronger reference answer that the platform generates with a more expensive model than miners are allowed to spend on. The platform mixes factual recall, explanation, comparison, and synthesis so memorized outputs do not win.
Miners implement a query entrypoint. Validators receive miner-task batches from the platform, run each script-on-task combination inside a sandbox, and score the response. According to the repo, scoring is total_score = comparison_score, where comparison_score is a pairwise judge against the reference answer, run twice with the order swapped to dampen position bias. Ties go to the run with lower total tool cost. Validators report scored runs back to the platform and submit weights on-chain.
Champion selection is the part worth understanding. The platform keeps an incumbent champion and walks through challengers in batch order. A challenger only takes the throne if it beats the incumbent by a sufficient score margin, or is effectively non-regressing while being materially better on runtime or cost. Inside the tolerance band, small score differences do not trigger a swap. The highest score in the batch is not automatically the new champion.
That design choice flows directly into emissions. Total miner weight is capped at 0.20 * latest champion batch score, and owner uid 0 receives the remainder, which burns. Today that produces a roughly 80% miner emission burn, by construction, not by neglect. Until a champion materially improves the benchmark, most of the miner share is destroyed rather than paid out. If no champion has been selected for a round, miner emission burns entirely.
The economics, then, are deliberately tight. With about 1,362 TAO in the pool, a market cap around 5,628 TAO, and roughly 1.53% of network emissions on a smoothed basis, the subnet is small but has been moving: roughly +22% on price over seven days and +34% over thirty, with about 272 TAO of net inflow over the last week. Active miner count is currently 1, which is consistent with a champion-driven design where one strong agent dominates until it gets dethroned.
- Execution and concentration: The repo is about two months old with three contributors, one of whom has written the large majority of the code. A small core team is normal at this stage but means key-person risk is real until the contributor base widens.
- Liquidity: With about 1,362 TAO of TAO-side depth in the pool and a market cap near 5,628 TAO, position sizing matters. Slippage on entry and exit can be material relative to fair value.
- Mechanism novelty: Champion-only emission with an 80% baseline burn is unusual. It is internally consistent with the "only reward genuine improvement" thesis, but it is also unproven at larger participant counts and could discourage challenger entry if the dethroning bar feels too high in practice.
- Competition: Search and research is a crowded centralized market and is starting to crowd up on Bittensor as well. Harnyx will be judged against both closed deep-research products and any neighboring subnets that target the same workflow.
- Audience build: The product is shipping; growing the validator set, the miner pool, and downstream users of the research output is the near-term gating factor on whether the design pays off.
Into the next one.