Research
Built on research. Tested against a working product.
AgenticObjects is grounded in established research on attention, memory, decision-making and trust in machine advice, then tested against the behaviour of a working product. This library brings together research papers, technical white papers and product studies: the science that shaped the architecture, the engineering decisions that follow from it, and the operating record we use to challenge our own assumptions. Read in any order. Copy what is useful.
AgenticObjects · Technical white paper · September 2026
From Answers to Records
A technical guide to agentic analytics you can stand behind.
The chat answer, the generated document, the emailed digest, the dashboard. Four containers, four different-looking failures, one shared cause: nothing that would let you defend the number travels with the number. This paper tests all four against a definition of “record” that predates the category by decades, sets out the five properties an AI output has to carry instead, and then does the part a demo never does. It opens one real scheduled run and shows what the publication gates held, what they let through, and why.
In one scheduled run at 05:00, the engine sent the agent’s entire draft answer back because fourteen figures in it had no source. Across the whole record, nine of the eleven held objects were the same sentence: fastest-growing, weakest, highest.
For CIOs, CTOs, data platform owners and the technical evaluator on a buying team.
Research paper · 2026
The Science Analytics Never Shipped
Forty years of cognitive science on attention, memory and trust, and why almost none of it reached an analytics default.
An agent can now produce findings continuously, at a cost approaching zero. The person receiving them has the same working memory they had in 1984, the last year a result from cognitive science made it into an analytics product. This paper is about the research that piled up in between: what explanations actually buy you, what they cost, and why the fix that feels obvious made things measurably worse in controlled studies.
“Explanations increased the chance that humans will accept the AI’s recommendation, regardless of its correctness.”
For CIOs, CDOs, heads of analytics, and anyone who has ever built a dashboard nobody opened.
Product vision
Not a paper: our thesis for agentic analytics, what runs today, the seven layers our agents are designed to read next and the order we build them in. Horizons, not dates.
Product studies
What we measured on a working product
AgenticObjects · Product study · October 2026
Which Model for Which Job
Eighteen language models, three jobs, one governed product.
We put eighteen models from five providers through three jobs the product does every day: writing a data catalogue’s descriptions, producing a CFO agent’s nightly briefing, and answering a finance question in chat. Two judges from different vendors scored every output without knowing which model wrote it. No model won all three jobs, models served from inside Türkiye came second in two of them, and the engine that checks every number removed 107 figures the models wrote but could not trace to a query result.
The quickest chat answers rarely said what to do next. The fastest models answered in 12 to 18 seconds on average, with reasoning off; their actionability averaged 1.6 to 2.1 out of 5, against 4.4 for the model that led. Which models those are is on the study page.
For CIOs, CTOs, heads of data and the technical evaluator choosing a model for governed analytics.
Also on this site, with nothing to fill in
Twelve questions for every vendor in this category
Open HTML. Copy it, put our name on nothing, ask them of us too.
See the checks hold on your own data
Four weeks on your warehouse, with the acceptance criteria written before week one. Or start smaller: a free data discovery study tells you which signals your semantic model carries today.