The Era of Ingestion – What Agents Cannot Hold

On 19 May 2026, from the I/O stage in Mountain View, Sundar Pichai, the CEO of Google, declared that the company he leads had entered its “agentic Gemini era.” The announcement was accompanied by a slate of products — Gemini Omni and Gemini 3.5, new agentic experiences in Search, an agent-first development platform called Antigravity, and Gemini Spark, a 24/7 AI agent that runs in the background on Google Cloud and continues working when the user closes their laptop. The Gemini app itself was reframed as more agentic, delivering proactive, 24/7 help. The word agentic appeared in nearly every product update on the keynote programme. The era had a name, a date, and a declared architect.

What an agent of this kind does, in the form Google has now made public, is precise enough to describe. Consider a researcher working on the descendants of the Romanov line. The work is slow, archival, multilingual, and inhabits a particular intersection of Russian dynastic history, post-revolutionary diaspora movements, and the recent emergence of DNA-based identification as a tool of last resort for disputed claims. The researcher could now configure a Spark agent to monitor news feeds, surface Russian-language archival announcements, flag auction lots containing letters or jewellery with documented provenance, track court filings in the European jurisdictions where Romanov-adjacent claims have surfaced over the past three decades, and route the most pressing items to the top of a daily digest. The agent would learn, with use, which signals matter to this particular researcher and which can be deprioritised. It would draft summaries from scattered notes. It would execute recurring searches. It would operate on cloud infrastructure independent of the researcher’s device, so that the work continues when the laptop is closed. This is genuine technical accomplishment, and the researcher’s days would be different for it.

What the agent cannot do begins where the description above leaves off. It cannot weigh whether a new descendant claim surfacing in a regional Russian newspaper is significant or trivial. It cannot distinguish a fourth-cousin-twice-removed asserting a tenuous link from a substantive claimant emerging with documented standing. It cannot place a finding within the broader inquiry the researcher is conducting — what the inquiry is for, whom it concerns, what its eventual disposition will be. It cannot understand that a finding might affect a living relative, complicate a discreet arrangement that has held for decades, or alter the texture of an inheritance dispute. It cannot exercise discretion about what should be surfaced openly and what should be held; Spark logs everything to Google Cloud, where holding is not, structurally, a concept it can perform. It cannot answer the question of what the finding means. It can only answer the question of what has appeared.

That this is so is not a defect of the engineering. It is a function of the architectural origin from which an agent of this kind is built. Spark exists because Google already owns the data position from which an agent can plausibly be assembled. Gmail, Calendar, Drive, Search, Chrome, YouTube, Maps, Android — the accumulation across two decades is what makes a 24/7 background agent technically and commercially feasible. The agent is the form an accumulated data layer takes once it begins to act on behalf of its users. Without that accumulation it cannot exist; with it, an agent is the natural next product. The cleverness of Spark sits in the harness rather than in the substance. The substance — the data — was already there.

This clarifies what the agent is, and by extension what it is not. Spark is downstream of mass data. It scales by adding users; it personalises at the edge, where it learns each user’s preferences and routines; it operates on behavioural patterns aggregated across nine hundred million Gemini users globally. It is personal in the sense that it customises around the individual. It is not bespoke in the sense of being trained on a single principal’s particulars — their formation, their family system, their holdings, their history, their register. The two words have come to be used interchangeably in the consumer market. The distinction is architectural and survives the consumer market’s collapse of it.

A different kind of intelligence is required for what sits above the agent’s reach. The Bespoke AI Confidante is downstream of the individual principal rather than of mass data. It is trained on the particulars of one person — what they have inherited, what they have built, what they collect, what they hold in discretion, what they are currently deciding. It does not surface what the user is most likely to need; it interprets what has been surfaced, in the context of who the principal actually is. It does not act on behalf of the principal; it thinks alongside them. The quality of an agent is a function of the breadth of the database it draws on. The quality of a Bespoke AI is a function of the depth of the formation it has absorbed. They are different categories of intelligence with different architectural origins, addressed to different orders of question.

The picture that emerges from the I/O 2026 announcement is therefore not two layers but three. At the base, the data platforms — genealogical databases, art market archives, family office reporting systems — that supply records and prices and ownership structures. Above them, now formalised by Google’s declaration, a layer of agents that ingest from those platforms, filter, route, digest, and increasingly execute on behalf of the user. Above the agents, the interpretive tier, where what has been surfaced is held against the question of what it means. The agentic era extends the second layer with considerable engineering accomplishment. It does not reach the third.

The Bespoke AI Confidante by SMA Crown Confidential occupies the third tier — not as a replacement for the AI agent, but as the layer above it, where interpretation, discretion, and the steady holding of context belong. Agents ingest; Bespoke AI Confidantes interpret. The agentic era, declared from the I/O stage on 19 May 2026, names the moment we now inhabit. What it cannot hold names the moment that has always required something else.

Seraphina Michelina Evelyna Bogomolova-Huotelin

Founder & CEO of SMA Crown Confidential

Digital Confidantes: Bespoke AI intelligence for private decision-makers