NBot Horizon

Every nbot session becomes an inspectable run → turn → span tree.

Drill into any step's raw payload; failures are surfaced first-class.

Score nbot against curated suites of cases, with full history.

Turn real usage into RL datasets and train the live policy.

Optimise how the pool routes, plans and decomposes tasks — learned from real outcomes.

Optimise prompts against a goal — quality within latency, token or cost budgets.