How a Context Layer Platform Makes AI Agents More Accurate and Less Expensive

Abstract

With Port of Context, a Context Layer Platform (CLP), Google’s Gemini 3.7 Flash model achieved 9.5 percentage points higher task success and used 64.7% fewer total tokens than the recorded public results for the same model on the airline customer-support benchmark τ²-Bench Airline. At identical uncached token prices, these airline results represent 69.3% lower estimated inference cost per successful task.

On the banking customer-support benchmark τ³-Banking, the model achieved 6.4 percentage points higher task success and used 37.2% fewer total tokens than its recorded public reference. For banking, the combination of higher task success and lower token usage yielded 46.3% lower estimated inference cost per successful task at identical uncached token prices.

We retained the benchmark task definitions and underlying business tools. Port of Context changed how the model accessed those tools and, in banking, how it located and retrieved policy evidence. In both benchmarks, the agent completed more tasks correctly while consuming fewer tokens, so accuracy and efficiency improved together.

The CLP gives an agent source routes, focused retrieval and typed tool execution through a shared gateway. We examine the completed conversations to explain how those capabilities supported the work. The public comparisons use separate runs with different simulators, harness settings and repetition counts; they do not isolate the contribution of each component.