Snowflake’s HybridResearcher shows why SQL and web-search agents need bridge training

Snowflake’s new HybridResearcher work isolates a failure mode that matters for enterprise data agents: a model can be competent at SQL and web search separately yet still fail when the output of one tool must become the input to the other.

The company starts with a representative task: identify suppliers affected by an external event, then calculate internal spend. That requires a governed database query, current web evidence and a reliable handoff between them. In Snowflake’s 120-question hard subset of HybridDeepResearch, simply fine-tuning Qwen3.5-4B on SQL-only, search-only or mixed single-source trajectories did not teach that handoff. SQL-only training reached 31.67% pass@8 versus 30.83% for the base model; the mixed set reached 29.17%. Snowflake Engineering

Train the dependency, not just the tools

HybridResearcher creates two task directions. In SQL-to-search tasks, a value returned by SQL becomes the entity the agent must investigate online. In search-to-SQL tasks, web research supplies a value needed in the SQL filter. The pipeline begins with validated text-to-SQL or search tasks, then synthesizes the missing source around a “bridge entity.” Deterministic checks reject examples where the bridge leaks into the prompt, where one source can answer alone or where removing the bridged SQL value does not change the result. Snowflake Engineering

That filtering matters because superficially successful traces were often shortcuts. Snowflake says 79.5% of initially passing teacher trajectories never executed a meaningful SQL query. Training on those examples would reward an agent for answering from memory while appearing to use both sources. Snowflake Engineering

The final recipe combines supervised fine-tuning with reinforcement learning. Snowflake generated teacher traces with larger Qwen and DeepSeek models, retained judge-approved trajectories and added existing hybrid and text-to-SQL examples. It then reduced 21,000 difficult reinforcement-learning tasks to 700 higher-signal items. The company reports that the curated subset produced useful learning signal in 71% of rollout groups, versus 38% for the full pool. Snowflake Engineering

Read the result with the right caveats

HybridResearcher-4B reached 60.83% overall pass@8, up from 30.83% for Qwen3.5-4B and 8.33 points above Qwen3.5-397B-A17B-FP8 in Snowflake’s setup. Its average accuracy across eight attempts was 27.19%, however, and SQL-to-search remained the weakest direction at 30% pass@8. These are vendor-run results on a 120-question hard subset, not an independent production evaluation. Snowflake Engineering

For practitioners, the lesson is narrower and more useful than “small models beat large models.” Build evaluations around cross-tool dependencies. Verify that the first tool actually produced the value consumed by the second, reject traces that skip required work and measure average reliability alongside best-of-eight success. A data agent’s hardest bug may live between two individually capable tools.