Execution patterns applied before inference: 84 (Source: Teradata company statement, reported by InfoWorld (Sept 28, 2026))
A cost layer bolted onto agentic AI
Teradata is adding three components — a Context Engine, an execution layer called Tera Harness, and reusable agent skills — to Tera, the AI workspace it introduced in May under its Autonomous Knowledge Platform. The company says the point is to stop agents from spending compute on steps that don't move a task forward, while still preserving enough business context to route work to the right data, tools, and models. Harness builds an execution plan before any call reaches a large language model, batches independent tasks together, and drops model or tool calls it judges won't advance the task. General availability is scheduled for December 2026.
The efficiency numbers Teradata is publishing
In its own evaluation on the SWE-bench Pro benchmark, Teradata compared Tera against Anthropic's Claude Code while both ran the identical Opus 5 model — isolating the test to the orchestration layer rather than the underlying model itself. The company reported lower token use, faster completion, and lower total cost, alongside higher task completion rates. The two figures are not interchangeable: the 73% reduction measures tokens consumed, while the 58% figure measures total dollar cost of the same tasks — a distinction worth holding onto before treating either number as a budget line item.
Why the number that worries CIOs isn't the model's list price
Analysts framed the real issue as unpredictability rather than raw price. Ashish Chaturvedi of HFS Research said agents left to reason freely will "happily burn tokens on loops that never move the task forward," and that even small reductions in calls per workflow compound at scale. Advait Patel, a senior site reliability engineer at Broadcom, said the harder problem for budget-holders is that agent costs today are nearly impossible to forecast, because in his words "the same task can take five calls one day and fifty the next." A harness that plans the route before execution starts, he said, makes that spend more consistent — which is what turns agentic AI into something finance can actually plan around.
The trade-offs a budget-holder should push on
Patel also pointed to a trade-off built into automated pruning: judging a step unnecessary is inherently subjective, and when Harness gets that judgment wrong, the enterprise pays less for a weaker answer rather than getting the intended savings — so outputs need closer checking, not less, once pruning is switched on. Robert Kramer, managing partner at KramerERP, was the one analyst to directly challenge how the efficiency numbers should be used: a vendor's own performance benchmark, he said, is not a stand-in for what an enterprise actually spends. "Teradata's benchmark results should not be treated as equivalent to enterprise total cost of ownership," he said. He argues CIOs should instead price out everything it takes to finish a business task — not just model usage, but the compute behind the data, the tool and retry calls, the orchestration layer, and the staff hours spent checking whether the output is right. Stephanie Walter of HyperFrame Research raised a separate, longer-term concern: the more a workflow depends on Tera's context, execution, and skills layer, the harder that workflow becomes to move to another platform later. Walter also said existing Teradata customers running complex, governed data environments are the most likely early adopters, since Tera would be "a logical expansion of an environment they already use" — while shops standardized on Snowflake or Databricks would be a harder sell because, as she put it, "the agents are not a good enough reason for lift and shift."
What to ask your team
Before treating Teradata's efficiency claims as a budget line item, ask engineering to reproduce the comparison against your own workflows and data volumes rather than SWE-bench Pro tasks. Ask finance to track cost per completed business task — including retries and human review time — instead of accepting per-token or per-run pricing as the whole story. Ask what the review process looks like when Harness prunes a step that turns out to have mattered, and who signs off on the result. And ask how much of your agent logic, skills, and context configuration would need to be rebuilt if you later moved off Tera — a lock-in question analysts say grows the longer a workflow depends on the platform.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: InfoWorld.
