Teradata has turned its assistant into what it calls an agentic coworker, and announced it with benchmarks against named competitors. In Teradata Transforms Tera into an Agentic Coworker, the company introduces three pieces: Tera Context Engine, a “vendor-neutral context and orchestration layer that gives AI governed business knowledge,” Tera Harness, “an intelligent execution layer that routes work across the right skills, tools, data, and models,” and Agent Skills purpose-built for data engineering, analysis and science. The framing line is “Where general-purpose AI assistants generate answers, Tera delivers outcomes.” All three ship in the fourth quarter of 2026.
The figures are the vendor’s own, from a press release with forward-looking statements attached, and they should be read that way. On SWE-bench Pro, using the same Opus 5 model, Teradata says Tera consumed 73 percent fewer tokens than Claude Code, completed work 42 percent faster and cost 58 percent less. On data-eng-bench it claims 53 percent lower cost per solved task than Snowflake Cortex Code, from Snowflake’s published data. Tera Harness is said to support 512 concurrent agents on a single 8-vCPU VM at 279 tool calls per minute, and to apply “84 proven execution patterns pre-inference.” The release says methodology and materials for independent validation are available, which is the right thing to say and the thing to check. What survives the marketing is the architecture: pre-inference guardrails, human approvals “embedded directly inside the agent loop,” checkpointed execution that can “pause for human approval at zero compute cost,” and MCP connectivity for customers’ own tools.
The catalog reads the platform underneath. The Teradata provider page lists 34 API pages, and the surface an agentic coworker would drive runs through the Teradata Queries API, the Teradata Sessions API, the Teradata Systems API and the Teradata Operations API that handles the platform work the release assigns to its Platform Agents. The agentic access profile maps 170 operations, 93 of them acting, with 3 marked human-in-the-loop. On the Agent Readiness score, agent skills, dry-run mode, reversibility, error semantics and examples are lit, which is a strong operational base for a harness that promises to block destructive actions.
The Kin Score is 59.4, strong band, carried by discoverability at 83.9 and developer ergonomics at 74.4, with access clarity at 60.5, contract quality at 56.1, contract governance at 31.8 and operational transparency at 21.1. The Agent Readiness score is 31.7, agent-ready, and the unlit list is where the release’s promises meet the public record. The MCP server is unlit, so the connectivity the release describes is not yet something the catalog can find. Delegated identity, consent identity, protected resource metadata and dynamic client registration are all unlit. A harness that embeds human approvals in the loop needs a way to say which human, acting under which grant. Teradata says Tera is “secured by enterprise identity and policy.” The record does not yet show how an agent arrives carrying that identity.