Teradata adds context engine and execution layer to cut AI agent token costs
Teradata is enhancing its Tera AI workspace with a context engine, an execution layer called Tera Harness, and reusable agent skills to streamline multistep data workflows. The new components aim to reduce unnecessary model and tool calls by planning execution before inference, batching independent tasks, and limiting unproductive loops. Teradata reports that on the SWE-bench Pro benchmark, Tera used 73% fewer tokens and completed tasks 42% faster than Claude Code while running the same model.
Teradata's Tera Harness applies 84 execution patterns before any inference occurs, batching independent tasks and discarding calls that don't advance the workflow. The company's benchmark testing on SWE-bench Pro, using the same Opus 5 model as Claude Code, showed a 73% token reduction, 42% faster completion, and 58% lower total cost. The platform also limits workflow steps based on progress to prevent unproductive agent loops.
The context engine supplies business guidance that routes each task to appropriate data, tools, models, and skills. Analysts note that planning-before-inference and task batching are established engineering practices, but embedding them as platform defaults could spare development teams from building and tuning these controls individually for every agent workflow.
This cost-control approach could reshape how enterprises budget for AI agents, making expenses more predictable as workloads scale. If token consumption becomes more consistent, organizations may feel more confident deploying agents across broader business functions. However, tighter execution constraints could limit agent flexibility in complex, unpredictable scenarios, and reliance on a single vendor's platform may raise lock-in concerns. CIOs will need to weigh efficiency gains against potential rigidity when deciding whether such harnesses suit their specific workflows.