Command centers for data stewards and context engineers

Every agent answers from the definitions it was last given. Ask one what net revenue was in EMEA last quarter and the answer depends on whether “net revenue” still means what the finance playbook said in March, which tables hold the bookings, and how they join. That knowledge lives, in most companies, in a Google Doc, a GitHub repo, a Confluence page and three people’s heads. The moment finance changes how partner credits are treated, every agent with a hard-coded formula starts answering wrong, silently and at once.

Google’s announcement names that failure mode and ships a loop against it.^[1] This piece explains what was shipped, what a context command center actually has to do, why two roles that used to work in two tools now share one screen, and what the pattern means for brand, marketing and GTM teams whose vocabulary — “qualified lead”, “active customer”, “campaign”, “SKU” — is the context agents reason over.

01 · What Google announced on 23 October 2026

The facts below come from Google Cloud’s announcement, dated 23 October 2026 and authored by Firat Elbey, Group Product Manager for Data Analytics, and Vasiya Krishnan, Product Lead.^[1]

• Self-learning context enrichment (preview) analyzes agent and user interaction traces — from Gemini agents, Conversational Analytics in BigQuery or custom agents, with PII redacted — to find where agents struggled: ambiguous column lookups, ungrounded metrics, missing definitions, users rephrasing the same question. It then finds the authoritative source and stages a fix.
• Knowledge Studio (preview) is the redesigned Knowledge Catalog experience for context engineers and data stewards: a visual map of the cross-cloud data estate, conversational exploration of the context graph, a review queue for proposed enrichments with source citations and a confidence score, and monitoring of enrichment runs including token usage and run duration.
• Context sets are governed subsets of the enterprise context graph — definitions, metrics, join logic, golden queries — scoped to a domain, team, agent or user group and controlled with IAM. The same catalog can hold a Sales definition of “net revenue” and a Finance definition without conflict.
• Human control is configurable. Auto-publishing can be switched on or off. Changes above a steward-set confidence threshold — Google’s example is 85% — are approved automatically; anything below it, and sensitive changes such as deletions, become a Change Request a steward approves, edits or rejects. Every change is logged with provenance and can be reverted.
• Claimed impact. Google states that Knowledge Catalog makes agents nearly 4× more accurate, with up to 33% lower latency and up to 75% fewer tokens, “based on internal benchmarks”. No method or external replication was published with the claim.
• Availability. The features “will be rolled out to customers in the next few weeks”, with a sign-up form for early access. Pricing was not stated.

Two pieces of lineage matter for anyone searching for this product. Knowledge Catalog is the name Dataplex Universal Catalog has carried since 10 April 2026; the API, client library, CLI and IAM names did not change.^[3] And the context sets inside it build on the Open Knowledge Format, which Google published as an open specification on 12 June 2026.^[2]

02 · What a context command center is

The market does not have a settled definition yet, so here is the one we use at auxfirst.

The word command is doing real work. A catalog is an inventory: it tells you what exists. A command center is an operations room: it tells you what is going wrong right now, what the system proposes to do about it, and gives a human the throttle. The shift Google describes — from “a repository of static inventory” to “an interactive command center”, in its own words^[1] — is the shift infrastructure made a decade ago, when monitoring dashboards turned into incident consoles with runbooks and on-call rotations.

03 · Why data stewards and context engineers now share one console

Until recently these were two jobs in two tools.

Data stewards own meaning: what a metric means, which table is the system of record, which join is correct, which definition is deprecated. Their tools were catalogs, glossaries and governance workflows built for humans reading documentation.

Context engineers own what an agent is given to reason with: system prompts, retrieval configuration, tool descriptions, schema snippets, golden queries, and the instruction files — AGENTS.md, CLAUDE.md, brand.md and their relatives — that shape agent behavior. Their tools were code repositories, prompt editors and evaluation harnesses.

Google’s announcement names the breaking point: context engineers writing prompts and mappings agent by agent, data stewards keeping definitions current across thousands of assets by hand, and neither approach surviving the move from a handful of agents to thousands.^[1]

04 · How self-learning enrichment works, step by step

Stripped of product language, the loop Google describes has six stages.^[1]

1. Observe. Interaction traces from Gemini agents, BigQuery Conversational Analytics or custom agents are analyzed, with PII redacted.
2. Detect. The system flags patterns that indicate missing or outdated context: ambiguous column lookups, metrics with no grounding, repeated rephrasing of the same question.
3. Source. It searches connected knowledge — Google Workspace, GitHub, Confluence, SharePoint, Cloud Storage, and first- and third-party structured sources — for the authoritative definition, and links unstructured sources such as playbooks to structured metadata such as tables and schemas.
4. Propose. It drafts business definitions, table and column descriptions, join relationships, golden queries and task-specific agent instructions, each with citations and a confidence score.
5. Gate. Proposals above the configured threshold publish automatically if auto-publish is on; everything else, and all deletions, becomes a Change Request for a steward. An organization can supply its own evaluation set to test proposals end to end.
6. Publish, log, revert. Approved changes land in a governed context set. Every change carries full provenance and can be rolled back.

05 · The gate, read on the Action Heat Ladder

Readers of our work on the Action Heat Ladder will recognize the structure of stage five. The confidence threshold is a heat setting: it decides how much autonomy the enrichment agent gets per class of change. Deletions are pinned to the hot end regardless of confidence — a hard gate, not a score. Google got that ordering right, and it is the ordering most teams get wrong when they first wire a model’s confidence into a workflow.

The second thing readers will recognize is the Change Request queue. It is the verification throughput constraint we described in the Publisher Agent Map: the number of outputs a qualified person can genuinely check in a day. Enrichment proposes faster than stewards can review, so the real capacity of the loop is the review bench, not the model. A team that raises its threshold to clear the queue is not getting more automation; it is getting less review.

06 · Context sets: scoping context per audience

A context set is the unit of publication. Google positions it as a curated, governed subset of the enterprise context graph for a specific domain, team, agent or group of end users, with IAM controlling who can read, use or edit it.^[1] A sales agent bound to the Sales set and a finance analyst’s assistant bound to the Finance set can each hold a legitimate, different definition of the same term. Agents are not locked to one set; through search in Knowledge Catalog an agent can locate the right set for a task.

Two properties matter for governance teams. Access to the underlying data still follows existing permissions: a context set tells an agent how to interpret data, it does not grant access to it. And portability: context sets build on the Open Knowledge Format, so an OKF bundle can be imported as a set and a set exported back out. Google states this explicitly as the way to keep context from being locked to its catalog.^[1]

07 · The Open Knowledge Format: the portable substrate underneath

The Open Knowledge Format is an open specification Google Cloud published on 12 June 2026, authored by Sam McVeety and Amir Hormati.^[2] It represents knowledge as a directory of Markdown files with YAML front matter. Each file is one concept — a table, a dataset, a metric, a playbook, a runbook, an API — and the file path is the concept’s identity. Concepts link to each other with ordinary Markdown links, which turns the directory into a graph. The only required front-matter field is type; title, description, resource and tags are optional. Optional index.md files support progressive disclosure as agents navigate; optional log.md files hold change history.^[4]

# a bundle is a directory; the unit of distribution
acme-retail/
├── index.md              # optional: directory listing for progressive disclosure
├── log.md                # optional: chronological history of updates
├── metrics/
│   ├── index.md
│   └── net-revenue.md    # one concept = one file; the path is its identity
├── tables/
│   └── orders.md
└── playbooks/
    └── partner-credits.md

# metrics/net-revenue.md
---
type: Metric                      # the only required field
title: Net revenue
description: Bookings net of discounts and partner credits
resource: bq://acme.finance.net_revenue_v3
tags: [finance, revenue]
---
Net revenue is computed from [orders](../tables/orders.md)
per the [partner credits playbook](../playbooks/partner-credits.md).

Google frames OKF as a formalization of the “LLM wiki” pattern — Obsidian vaults wired to coding agents, the AGENTS.md and CLAUDE.md family of convention files, repositories full of index.md and log.md artifacts, metadata-as-code inside data teams — so that wikis written by different producers can be consumed by different agents without translation.^[2]

Why this matters to the command-center argument: the console is Google’s, but the substrate is files. If a steward’s approved definitions can leave as a directory of Markdown that any other tool can read, the command center is a control plane over knowledge the organization still owns. That is the condition under which it is safe to let a vendor’s console become the place where business truth is edited. We have argued the same point from the other direction in Markdown for AI and in the template for AGENTS.md for teams that don’t write code: business context belongs in plain, versioned, human-readable files, and the tooling around them should be swappable.

08 · What this means for brand, marketing and GTM teams

The example Google uses is a finance question. Swap in your own vocabulary and the stakes are identical.

09 · What is known, what is claimed, what is unknown

↔ scroll table

The AUX takeaway

The interesting word in Google’s announcement is not self-learning. It is command. A system that proposes changes to what your agents believe, at machine speed, needs a room with a human in it, a throttle that human controls, and a log that shows what was changed and by whom. Google has built the room. What it cannot build for you is the roster: who sits in it, which classes of change they own, and what happens when the queue is longer than their day.

A catalog tells you what exists. A command center tells you what is going wrong, what the system proposes, and who gets to say yes.

Business context was always an operated system. It was just operated badly, in decks and heads. The announcement makes the operations visible, which is the first step to making them accountable — and that, not the 4×, is the part worth staffing for.

Terms used here

Context command center

auxfirst's term for the operating console where the people accountable for an organization's business definitions can see what agents currently know, see where that knowledge fails in live use, review machine-proposed corrections against their sources, decide what is published to which agents, and roll back anything that goes wrong, with every decision logged.

Self-learning context enrichment

A Knowledge Catalog capability, in preview from 23 October 2026, that analyzes agent and user interaction traces to find missing or outdated business context, finds the authoritative definition in connected documents, repositories and tables, and proposes a fix with citations and a confidence score.

Context set

A curated, governed subset of the enterprise context graph, holding the definitions, metrics, join logic and golden queries for one domain, team, agent or user group, with IAM deciding who can read, use or edit it. Context sets build on the Open Knowledge Format and can be imported from and exported to it.

Context engineer

The role that decides what an AI agent is given to reason with: definitions, schema context, tool descriptions, golden queries and instruction files, and that tests whether that context produces correct behavior.

Data steward

The role that owns meaning: what a metric means, which table is the system of record, which join is correct, which definition is deprecated. In a command center, the steward's approval is what changes the agent's behavior.

Open Knowledge Format (OKF)

An open specification from Google Cloud, first published 12 June 2026 and at version 0.2 by October, that represents business knowledge as a directory of Markdown files with YAML front matter. Each file is one concept, links between files form a graph, and only a type field is required.

Change Request

In Knowledge Catalog, a proposed context change that did not clear the steward-set confidence threshold, or that is sensitive such as a deletion, and therefore waits for a steward to approve, edit or reject it.

Verification throughput

auxfirst's term for the number of agent outputs a qualified person can genuinely check in a day. In an enrichment loop it is the real capacity of the system, because proposals arrive at machine speed and approvals happen at human speed.

Questions people are asking this week

Method · Every source was fetched and read on 23 October 2026. Product facts come from Google’s two announcements and its documentation; nothing is claimed from the demo video, which is cited only as the demo Google itself links. Vendor performance figures are labeled as vendor figures wherever they appear. The five-capability test, the row split in Figure 4 and the two definitions of net revenue in Figure 5 are auxfirst’s, not Google’s. Product details reflect the announcement and will be revised when the features reach general availability.