SOMA

AI agents repeatedly send large amounts of context to language models—conversation history, tool outputs, repository files, and previous actions. As tasks grow longer and more complex, this context balloons, driving up inference costs and limiting agent efficiency. A single agent task can require dozens or hundreds of model calls, each resending significant portions of context.

SOMA addresses this by creating an open competition for developing better compression techniques. Rather than relying on a single approach, the network incentivizes a diverse set of miners to build compression algorithms, models, and hybrid strategies. Each submission is evaluated on real agent workloads, not static benchmarks, ensuring that compression actually preserves the agent's ability to complete tasks while reducing token consumption.

On Bittensor's netuid 114, miners submit compression solutions that are automatically tested against uncompressed baselines by validators. Scoring combines two signals: agent task performance (whether the agent still succeeds) and weighted token savings (distinguishing between input, cached input, and output tokens). Validators execute tasks multiple times to reduce variance, then report results to the platform. The incentive mechanism distributes rewards across three competition benchmarks—swebench_verified, swe_explorer_explore, and swe_explorer_edit—using a layered weighting system that rewards consistency, bonus performance on hard tasks, and penalties for quality regression. Top performers gain visibility and integration into production compression infrastructure.

SOMA is a Bittensor subnet focused on context compression for AI agents. Miners develop compression algorithms, models, and hybrid approaches to reduce token consumption while preserving agent task performance. Solutions are evaluated against uncompressed baselines on real agent workloads, with scoring based on both task quality and weighted token consumption.

The strongest approaches contribute to SOMA's production compression infrastructure.