Teradata (NYSE: TDC) is expanding Tera into an agentic coworker for enterprise data work, adding a context layer, an execution layer and reusable agent skills that it says will become available in the fourth quarter of 2026. The practical target is the operational cost of moving from an AI response to a governed action across data engineering, analytics and platform work.
The company said the new Tera Context Engine, Tera Harness and Agent Skills can be used together or separately within its Autonomous Knowledge Platform. Teradata is positioning the components for organizations whose data, models and workflows span cloud, on-premises and sovereign environments, and that need to retain control over where workloads run and how enterprise data is accessed.
Tera Context Engine is designed as a vendor-neutral context and orchestration layer. According to the company, it can connect databases, structured and unstructured data platforms, pipeline engines, catalogs, models and AI agents without moving data or requiring a single-vendor data estate. It is intended to connect metadata, lineage, business definitions and access policies so that an AI system has more than a raw prompt or a disconnected document store to work from.
That distinction matters for enterprise teams. An agent asked to build a pipeline specification or validate a data product needs the organization’s terms, policies and source lineage—not only a general-purpose model’s interpretation of a question. Teradata said its context engine will let agents apply governed context to data-product creation, pipeline specifications and validation controls, while retaining provenance that customers can inspect.
The second component, Tera Harness, is the execution runtime. The company said it coordinates skills, tools, data and models across multi-step work, applies guardrails and human approvals before execution, and maintains state so tasks can pause and resume after a failure. Teradata said the system uses pre-inference execution patterns to reduce unnecessary model and tool interactions rather than relying on repeated large-language-model reasoning for every step.
Teradata also described a Go-native and gRPC-based performance core for the runtime. In company testing, it said Tera supported 512 concurrent agents on a single eight-vCPU virtual machine while serving 279 tool calls per minute. Those are vendor-reported results rather than an independent benchmark, but they identify the design tradeoff Teradata is trying to address: agentic systems can become costly when every task launches separate processes and repeatedly invokes models.
The company said Tera includes platform agents for workload tuning, compute sizing, telemetry and FinOps, along with analytics agents for work such as natural-language-to-SQL, Python and query optimization. Its Agent Skills are intended to package task-specific guidance so advanced data and platform work does not require every user to have deep Teradata specialization. Customers will still need to assess the controls, integrations and model choices that fit their own environment.
Teradata reported that, in a comparison using the same Opus 5 model on SWE-bench Pro, Tera used 73% fewer tokens than Claude Code, completed work 42% faster and had 58% lower total cost. The company also cited results from data-eng-bench and ADE-bench. Those comparisons should be read with the published methodology and workload fit in mind, but they reinforce that token use and durable orchestration are now infrastructure concerns, not just model-selection questions.
For enterprise technology leaders, the new product plan shifts Teradata’s position from data-platform provider toward a control plane for AI work that must carry business context into execution. Compared with agent frameworks that leave context, governance and resilience to separate layers, the planned Tera stack aims to make those controls part of the data-work runtime itself.