On 22 September 2026, Teradata announced a major expansion of its Tera AI assistant: a Context Engine, an execution system called Tera Harness, and a catalog of Agent Skills designed so governed agents can carry out data tasks across enterprise systems. SiliconANGLE coverage places the move in the race to turn analytic copilots into agentic coworkers with real controls.
Tera is an agentic workspace and natural-language interface that lets analysts, data engineers and database administrators request analysis, pipelines or infrastructure management without writing SQL or code from scratch, while staying inside organizational access rules. Although it is part of Teradata’s Autonomous Knowledge Platform, the company says it can reach data held outside Teradata systems.
The new Tera Context Engine connects information from databases, catalogs, pipelines and other sources without forcing companies to move their data. It brings together metadata, lineage, business definitions and access policies so an agent can interpret each request in that organization’s context, and claims it can trace AI outputs back to their sources while applying the same policies as information moves between systems.
Tera Harness handles execution: it selects the tools, models and data needed for a task, tracks progress across multiple steps and can pause for human approval before sensitive actions. Teradata says the system can resume work after an infrastructure failure and applies controls before an action runs, precisely to limit unauthorized or destructive operations.
The third addition, Agent Skills, packages common data-engineering, analysis and data-science tasks into reusable functions. Agents can use the skills to write SQL or Python, optimize queries, tune workloads and size compute. Organizations can connect their own tools through the Model Context Protocol (MCP), the same standard gaining traction in enterprise agent hubs.
CPO Sumeet Arora framed the market problem: most enterprises are not starting from scratch with AI; they are dealing with tools that do not work together and a skills gap that blocks scale. Tera tries to close that gap by offering governed context plus execution with human brakes, not merely a chat over a warehouse.
Teradata also pushed efficiency numbers. In company-reported testing on SWE-bench Pro using the same Opus 5 model, Tera used 73% fewer tokens than Claude Code, completed tasks 42% faster and had 58% lower total cost, with a higher completion rate. On the data-eng-bench pipeline benchmark, it reported 53% lower cost per reliably solved task than Snowflake’s Cortex Code, based on Snowflake’s published results. The company cautioned that comparisons reflect its test conditions, not every enterprise workload.
The new capabilities, along with specialized skills, are planned for the fourth quarter of 2026. Customers will be able to choose models and run workloads in cloud, on-premises or sovereign environments. Teradata is also offering AI Services to help identify use cases, configure Industry Knowledge Models and move enterprise knowledge into production faster.
The announcement lands the same day SAP and Darktrace reinforced their agent-governance narratives, suggesting a coordinated enterprise-ecosystem wave: inventory agents, give them context and measure their cost. For data teams, Agent Skills is the most tangible piece because it turns scattered runbooks into callable contracts.
Risks are familiar for benchmark claims: token counts and costs depend on prompt, harness and dataset; 73% fewer tokens does not guarantee better analytic quality on proprietary pipelines. Still, a classic data-warehouse vendor publishing comparisons against Claude Code and Snowflake marks a shift in competitive terrain.
For compliance, Context Engine and Harness matter more than speed marketing. Tracing outputs to sources, applying policies in motion and pausing sensitive actions are controls that align with AI Act inventories and NIST frameworks that governance assistants are already embedding in agent hubs.
Neural Pulse sees convergence of three layers here: enterprise knowledge (context), multi-step execution (harness) and packaged abilities (skills). Without all three, a data agent either hallucinates or remains a SQL chatbot.
In short, on 22 September 2026 Teradata repositioned Tera as a governed agentic coworker, not a simple NL2SQL assistant. The real test arrives in Q4, when customers measure whether Context Engine and Agent Skills truly cut access incidents and token cost in production.
We will track Q4 availability, the first independent checks of claims versus Claude Code and Cortex, and MCP adoption as a bridge between Teradata skills and third-party tools.
