Arbos Is Testing a New Way to Evaluate Bittensor Subnets

Bittensor has no shortage of opinions on which subnets are building something valuable. The harder question is how to evaluate them objectively at scale.

Arbos is now testing one possible answer.

Arbos is an autonomous AI agent built by Bittensor co-founder Jacob Steeves (Const). It already operates a root validator, but has recently taken on another role: evaluating Bittensor’s subnets using a standardized scoring framework.

The agent looks at factors including whether a subnet provides a genuine product or service, whether it is actively operating, how its incentives are structured, whether its scoring can be independently verified, and the activity of its team and product.

Each subnet receives an IN or OUT assessment alongside a score out of 10.

The first major sweep is now complete, with Arbos having evaluated all 128 subnets. The results have already sparked discussion across the Bittensor community, particularly because some established subnets scored poorly while relatively unknown projects received high scores.

But there is an important caveat: these scores are currently a test. The Bittensor team has explicitly said they should not be treated as official rankings or investment signals. The purpose right now is to test whether an AI agent can reliably evaluate subnets, not to determine who receives emissions.

Why it could become important

The experiment becomes much more interesting if it works.

Bittensor’s subnet ecosystem is too large for any individual to continuously research every project, inspect its code, follow its incentive mechanisms, and determine whether its product delivers value. An autonomous evaluator could potentially do this work continuously and at scale.

The longer-term vision is even more significant. If Arbos can develop into a reliable and transparent evaluator, its assessments could eventually help inform root-validator allocation decisions, especially in the root reborn update, giving validators an intelligent way to decide which subnets deserve weight allocation.

In other words, Arbos is testing whether AI can move beyond building and scoring intelligence on Bittensor to also helping determine which intelligence is worth rewarding.

For now, the scores should be treated as experimental. But the underlying idea could become an important piece of Bittensor’s future if the evaluation system proves accurate, transparent, and resistant to bias.

Read more about this: