Teradata turns Tera into an agentic coworker for enterprise data

Teradata announced a major evolution of Tera into an agentic coworker for enterprise data work, introducing Tera Context Engine for governed business knowledge, Tera Harness for intelligent execution, and purpose-built Agent Skills for data engineering, analysis, and science.

SOFTWARE desk — enterprise data agents only matter if they carry governed business context into the execution loop, not just chat answers.

What the company says the product is for. Where a general-purpose assistant writes an answer and stops, the release says Tera is meant to deliver an outcome. Through ordinary language and guided steps, a business analyst, a platform engineer, or a database administrator can analyze data, build an AI application, operate infrastructure, and automate a workflow. The release says they do that from one governed environment that can reach enterprise data on more than one platform, not only inside Teradata. Industry knowledge is embedded in each interaction and locked to the company’s identity and access rules, so the person does not need to be a Teradata specialist. Governed, here, means those identity rules and access policies stay attached to the work. File that framing as the release’s.

Tera Context Engine, in the release’s words. The opening list calls it a vendor-neutral context and orchestration layer that gives AI governed business knowledge. A later section calls it an open, neutral layer that sits above the data systems a company already uses. It connects databases, structured and unstructured data platforms, pipeline engines, catalogs, models, and AI agents without moving the data, without ripping out the systems already installed, and without locking the company to one vendor. The release says the engine is not limited to data Teradata manages. A context layer is the shared picture of what the business means: names, relationships, and rules. Orchestration means coordinating those systems so an agent can use them together. Vendor-neutral means the layer is meant to work with more than one maker’s tools. Without moving data means the rows stay where they already live. File that description as the company’s. This desk did not connect a catalog.

What the engine adds, still the release. A native context graph turns scattered enterprise data into knowledge a later job can reuse. It connects metadata, lineage, semantics, and business meaning as relationships, and it keeps that knowledge tied to governance, lineage, and access controls as it moves across systems and agents. Metadata is the description of the data. Lineage is the record of where a number came from. Semantics is what a field means in the business, not only its column name. Industry Knowledge Models, which the release calls the product of Teradata’s human-checked expertise, give an agent a starting map of an industry’s terms, relationships, policies, and working conditions, so the agent does not have to invent that meaning from scratch. The release also says governed context can move work off repeated model calls and onto paths the company has already defined, and that agents can use that context to draft data products, pipeline specifications, and validation checks. Those are the company’s claims. This desk did not inspect a graph.

Tera Harness. The release calls it an intelligent execution layer that routes work across the right skills, tools, data, and models, and that keeps context across a workflow without a person choosing each hop. It says a general-purpose agent writes a response and leaves the person to act, while the harness is meant to turn an intention into a finished job. Before a step runs, the release says the harness applies guardrails, loads the memory that matters, and plans the work. A guardrail is a check that blocks a risky action. Safety rules, policy controls, and a person’s approval sit inside the agent loop, so a high-risk action is stopped before it runs. A loop, here, is the cycle of plan, act, and check. The release says that is different from asking the model to police itself, or from a filter that sits outside the loop. It also says the harness limits wasted repeats before they become a bill: it groups independent work, drops model and tool calls that do not change the outcome, and bounds a run by how far the task has gotten. It says Tera applies 84 proven execution patterns before the next model call, so the agent starts from a plan. 84 is the company’s count. This desk did not list the patterns.

A scale line, and whose number it is. The release says the harness supported 512 agents at the same time on one virtual machine with 8 virtual processors, and served 279 tool calls a minute. A virtual processor, shortened to vCPU, is a slice of a computer a cloud vendor rents. At the same time means those agents are running together, not one after another. 512 and 279 are Teradata’s figures. This desk did not load that machine. The release also says a job can run for days or weeks, pause for a person’s approval at no compute cost, and resume after a failure. Compute cost, here, means the bill for the machine while it waits. File those lines as the company’s. How the engine is written stays in Sources.

Agent Skills. The release says they are purpose-built for data engineering, data analysis, and data science. They are reusable capabilities for a specific task, packaging Teradata knowledge into functions an AI can call. A person can also ask for them in ordinary language. The release says skills load based on the task, and Tera routes each one to the models and tools it says fit. MCP connectivity lets a company add its own tools. MCP is the Model Context Protocol, the plug an agent uses to call a tool. The release names two kinds of Tera agents that draw on these skills. Platform Agents handle the operational work of the Teradata environment: workload tuning, sizing the compute, telemetry, and FinOps. Telemetry is the stream of measurements about how the system is running. FinOps is the work of watching what that computing costs. Analytics Agents do the data work: ordinary language into SQL and Python, and tightening a query. SQL is the language a database uses for a question. Python is a programming language used for analysis. File the list as the release’s. A skill name is not a demo this desk ran.

Where it sits, and what the customer keeps. The release says Tera is part of the Teradata Autonomous Knowledge Platform. The three capabilities can be combined or used on their own. A customer keeps the choice of which models to use, where the work runs, and how enterprise data is reached, across a cloud, on-premises computers, and sovereign environments. On-premises means the computers the customer owns. Sovereign, here, means an environment kept under a country’s or a company’s own control. The release also says Tera can run analytic and machine-learning work inside the Teradata environment, including the Teradata Console for database administrators, directly on enterprise data, without sending that work out to an outside model. It says that avoids moving the data and lowers the chance of a made-up number on jobs such as a forecast, a regression, or a segmentation. A forecast is a prediction of a future number. A regression fits a number to the factors that move it. A segmentation splits records into groups. Those lines are the company’s. This desk did not run a forecast.

The benchmark claims, and the limit on them. The release says that on SWE-bench Pro, using the same Opus 5 model, Tera used 73% fewer tokens than Claude Code, finished with a higher task-completion rate, completed the work 42% faster, and incurred 58% lower total cost. SWE-bench Pro is a public test of whether a coding agent can finish real software changes. A token is a small chunk of text a model reads or writes. 73% fewer tokens means Tera used a bit more than a quarter of the text Claude Code used, on Teradata’s account. 42% faster means the work took a bit more than half as long. 58% lower total cost means the bill was a bit under half. The release says the completion rate was higher. It does not print the two rates. Do not invent them. On data-eng-bench, which the release says Snowflake Labs and Bespoke Labs built for data-pipeline engineering, Teradata says Tera delivered 53% lower cost per reliably solved task than Snowflake Cortex Code, using Opus 5, based on published benchmark data. 53% lower means the cost for each task that counted as solved was a bit under half of that comparison, on Teradata’s claim. The release also says that across data-eng-bench and ADE-bench, Tera posted the highest Pass3 score on data-eng-bench and tied for the top score on ADE-bench. These figures are the company’s. They are not an outside lab’s result. This desk did not run the tasks.

What the company’s own write-up adds. The release points to a Teradata article, “Introducing Tera—the World’s Most Efficient Data Coworker,” by Anuj Agarwal, dated September 22, 2026, for the method and for materials it says an outsider could use to check the work. “World’s most efficient” is the article’s headline. This desk did not rank agents. The article says data-eng-bench has 103 tasks, a task passes only when every test check passes, and Pass3 means all three attempts succeed. It prints $1.23 per reliably delivered pipeline for Tera, $2.61 for Snowflake’s agent, which the article calls CoCo, and $3.40 for Claude Code. $1.23 is about 53 percent less than $2.61, which is the press release’s 53% line written as dollars. The article says SWE-bench Pro, from Scale AI, has 731 tasks, and that it is the benchmark they did not design Tera to target. It says both agents on that comparison drove the model it calls claude-opus-5. The press release’s name for that model is Opus 5. The article is still Teradata writing about Teradata. This desk did not rerun the 103 tasks or the 731.

When it is supposed to ship, and the quote. The release says Tera Context Engine, Tera Harness, and Agent Skills will be available in Q4 2026. Q4 is the fourth quarter, October through December 2026. That is a plan on the page. It is not a product this desk can open today. The page labels future availability, capabilities, and expected benefits as forward-looking statements. Sumeet Arora, chief product officer, said most enterprises are not starting from scratch. They have tools that do not work together, and a skills gap that makes those tools hard to use at a large scale. He said Tera is meant to put business context, intelligent execution, and ready-made expertise in the hands of every person who works with data. He said what those enterprises do not give up is control over their models, their data, and where everything runs. File the name, the title, and those sentences as the release’s. A quote is not a measured saving, and it is not a ship date that has arrived.

A services line, and not the lead. The release says Teradata AI Services is for organizations that want help getting to a production deployment. The services are described as finding the jobs where governed context should matter, setting up Industry Knowledge Models, and getting that knowledge into production faster. A production deployment means the software is doing real work, not a trial. That offer sits on the same page as the three capabilities. It is not a separate product launch, and this desk did not buy a workshop.

Plain English for the rest of the card: agentic coworker = software that takes steps on a data job with the company’s rules attached, not only a chat reply. Tera Context Engine = the layer that holds governed business knowledge and connects the systems an agent needs without moving the data. Tera Harness = the layer that routes each step across skills, tools, data, and models, with checks inside the loop. Agent Skills = ready-made capabilities for data engineering, data analysis, and data science. governed = identity and access rules stay with the work. vendor-neutral = meant to work with more than one maker. context = the names, relationships, and rules the business already uses. lineage = where a number came from. MCP = the plug an agent uses to call a tool. vCPU = a rented slice of a processor. 512 and 279 = the company’s scale line for one 8-vCPU machine. 84 = the company’s count of execution patterns. It is not a list this desk checked. token = a small chunk of text a model reads or writes. 73%, 42%, and 58% = Teradata’s SWE-bench Pro comparison with Claude Code on Opus 5. 53% = Teradata’s data-eng-bench cost comparison with Snowflake Cortex Code. Pass3 = all three tries succeed, on the company’s write-up. Q4 2026 = October through December 2026, a planned window, not a launch this desk watched. Opus 5 = the model name on the press release. The write-up’s name for the same comparison is claude-opus-5. This filing is the 22 Sep announcement.

PRIMARY here: Teradata’s 22 Sep 2026 press release, “Teradata Transforms Tera into an Agentic Coworker; Introduces Tera Context Engine and Tera Harness,” at teradata.com/press-releases/2026/tera-an-agentic-coworker — Tier A PRIMARY, the company’s own record. The visible dateline is Sep 22, 2026, San Diego. The page did not print an hour. The same-day Teradata article by Anuj Agarwal is the company’s benchmark write-up, not an outside lab and not a substitute for the release. The three capabilities, the vendor-neutral context layer, the harness routing and in-loop guardrails, the 84 patterns, the 512-agent and 279-call line, the Agent Skills split, the platform and analytics agents, the Q4 2026 availability, the Opus 5 comparisons of 73% fewer tokens, higher completion, 42% faster, and 58% lower total cost versus Claude Code, the 53% data-eng-bench cost line versus Snowflake Cortex Code, the Pass3 and ADE-bench lines, and the Arora title and quote are the release’s. The 103 tasks, the Pass3 definition, the $1.23, $2.61, and $3.40 figures, the 731-task count, and the claude-opus-5 name are the article’s. NOT claimed: that this desk ran Tera, reran SWE-bench Pro, data-eng-bench, or ADE-bench, a completion rate the release did not print, an independent ranking, that Q4 2026 has arrived, a price for the product, a stock tip, or investment advice. Distinct from the already-filed salesforce-trusted-enterprise-ai-harness, creatio-bank-ai-300m, and firecrawl-alexandria-75m.