TrajectoryRL

Deploying AI agents that use multiple open-weight models requires a routing decision at every request: which model to call, in what combination (serial, parallel, fallback), and what to send them. A static model choice (pinning one model) leaves performance on the table when different scenarios reward different trade-offs. A hand-coded decision tree is brittle and scenario-specific. What practitioners need is a systematic way to author, test, and refine these routing policies in an environment that surfaces their real cost and quality trade-offs.

TrajectoryRL Season 2 makes policy optimization a measurable competition. Miners submit fusion policies—programs that observe agent requests in real time and decide which models answer—evaluated against a benchmark of 26 real-world scenarios with hidden tests and deterministic verifiers. Each scenario runs in an isolated sandbox where the policy sees only the conversation stream (no ability to read task files or peek at tests), and can call any of eight allowlisted Engy models, up to a $1 per-scenario cap. The verifier runs fresh after the agent completes, so the policy cannot fake results. Validators evaluate independently and stake-weighted Bittensor consensus picks a winner; the winning policy is served as Engy's auto mode, available to every customer and harness.

The platform handles pack hosting, pre-eval (LLM-as-judge checks for hardcoding and scenario-specific tricks), metered LLM calls with cost tracking, and verifier orchestration. Miners author policies with no uptime requirement and no GPU—all evaluation happens on validators' hardware. Consensus is on-chain; the seat changes hands only when a challenger demonstrably outperforms the incumbent by a noise-aware margin.

TrajectoryRL (Subnet 11) is a Bittensor competition where miners submit fusion policies that decide which open-weight Engy models to call, in what combination, to serve an agent (Hermes). Validators evaluate each policy in isolated sandbox scenarios with hidden tests and deterministic verifiers. The seated winner's policy is served as Engy's auto mode.

Season 2 evaluates policies; Season 1 evaluated static instruction files (SKILL.md). Competition metric: sum of per-scenario tests passed across 26 scenarios. No GPU or uptime required from miners.