Behind Genie One's Accuracy: Why Curated Agents Need Governed Context

Genie One's accuracy problem is fundamentally a lack of context governance, not insufficient model capability.
Using Genie One as an example, the article explores why curated Agents' accuracy collapses in cross-departmental scenarios. Finance and underwriting give opposite answers to the same question—not due to data fabrication, but because each operates in a different data context. Governed context includes three indispensable layers: definition consistency, permission visibility, and version timeliness. Curated Agents excel in single-department evaluations but default to familiar conventions across domains and don't question their context, amplifying governance defects. The article proposes four engineering directions: unified semantic layer, context traveling with queries, output with lineage, and forced alignment for cross-domain questions, noting that Agent competition is shifting from model capability to context engineering quality, and that "precision" should be redefined as being reproducibly correct within governed context.
When Two Departments Give Opposite Answers
Ask the finance department and the underwriting department the same question—whether the loss ratio improved from 2024 to 2025—and you'll likely get two contradictory answers. The issue isn't that either department is lying; it's that each is speaking from a different data context.
This is precisely the core challenge facing Curated Agents like Genie One: the Agent itself can be very smart, but if the context fed to it hasn't been governed, its "precision" is only locally correct.

What Is "Governed Context"
Governed Context doesn't simply mean cleaning up the data. It encompasses at least three layers:
- Definition consistency: Is the loss ratio the same formula under finance and underwriting conventions? Which line items are included in the numerator and denominator?
- Permissions and visibility: Different roles can see different data scopes. The Agent needs to know "who is asking" in order to decide "what to give."
- Version and timeliness: Which snapshot the data comes from and which settlement cycle it postdates directly determines whether the conclusion is reproducible.
Without any one of these layers, the Agent may "correctly" give the wrong answer within its own context. Genie One's accuracy problem is fundamentally not a model capability problem but a context governance problem.
Semantic Layer is the technical vehicle for governed context, typically sitting between the data warehouse and consumption endpoints (BI, Agents), replacing scattered SQL with centralized metric definitions, dimension relationships, and computation logic. Implementations such as dbt Semantic Layer, LookML, or AtScale allow defining a single formula for metrics like "loss ratio" and exposing consistent query results to different frontends via API. The value of the semantic layer is that when conventions need to change, you modify the definition in one place, and all downstream consumers (including Agents) automatically inherit the new convention. For Agents, the semantic layer provides not only the correct numbers but also metric metadata (such as numerator/denominator line items, filter conditions, applicable time ranges), enabling the Agent to cite authoritative definitions in its answers rather than relying on its own vague understanding of the question.
The Accuracy Trap of Curated Agents
"Curated" means the Agent is confined to work within a specific domain, specific datasets, and specific tool sets. This constraint improves performance in specialized scenarios but also introduces new traps.
Local Optimum Is Not Global Correctness
An Agent confined to underwriting data will be highly sensitive to loss ratio changes under underwriting conventions, answering quickly and accurately. But when a question crosses departmental boundaries, it defaults to the definition it knows best. Accuracy metrics look great in single-department evaluations but collapse in cross-department scenarios.
Data Lineage describes the complete flow path of data from source to final consumption artifact, including table-level, column-level, and even row-level dependencies. In Agent scenarios, lineage is especially critical: when an Agent generates a conclusion, lineage information can tell users which ETL transformations the number went through, which dimension tables it was joined with, and which data snapshot it used. Without lineage, the Agent's output is a black box; with lineage, users can audit the Agent's reasoning basis just as they audit BI reports. Lineage collection typically relies on parsing SQL, ETL job logs, or dedicated metadata management tools (such as Apache Atlas, DataHub). For cross-departmental convention conflicts, lineage also helps pinpoint the root cause of discrepancies—whether it's different source systems, different transformation logic, or different snapshot times.
Governance Gaps Get Amplified by Agents
During manual analysis, an analyst will ask "how was this number calculated?" upon spotting a problem. An Agent won't proactively question its own context; it will pass governance defects straight through to the final answer—and in a more confident tone.
Accuracy Demands Greater Explainability
The accuracy of traditional BI reports can be audited—click on a number and you can see its lineage. An Agent's output is natural language; without binding to governed context, users find it hard to judge the data sources and computation logic behind the statement.
Reproducibility in the context of data analysis means that given the same inputs (question, context, data snapshot), the same output can be reliably obtained. For traditional reports, reproducibility is guaranteed by fixed SQL and snapshots; for Agents, additional challenges arise: the non-determinism of large models, changes in retrieval context, and even minor adjustments to prompt templates can all alter the output. Governed context provides a reproducible anchor for Agent output by fixing metric definitions, binding data snapshot versions, and recording permission and convention parameters at query time. A reproducible Agent output means: when business stakeholders dispute a conclusion, they can trace back to the specific context combination and determine whether the disagreement stems from the data itself or from the Agent's interpretation.
How to Build a Governable Context Layer for Agents
Drawing from the Genie One case, several engineering directions can be distilled for practical implementation.
Unified semantic layer first. Establish a metric definition hub before deploying the Agent, so that core metrics like "loss ratio" have only one authoritative definition. Convention differences across departments should be registered as explicit variants rather than scattered across their respective SQL.
Context travels with the query. The Agent should receive not just the question text but also the asker's identity, department, data permissions, and metric convention preferences. This information determines which context it should retrieve.
Output includes lineage. Every conclusion should be traceable to specific metric definitions, data snapshots, and computation paths. This is not just a compliance requirement but a prerequisite for building user trust.
Cross-domain questions force alignment. When a question spans multiple business domains, the Agent should align conventions first before answering, explicitly presenting different conclusions under different conventions when necessary—rather than quietly picking one.
Implications for the Industry
The Genie One case points to a broader trend: the competitive focus for Agents is shifting from model capability to context engineering quality. An Agent with access to all enterprise data but no governance layer poses greater risk than a less capable Agent with clean context.
For enterprises, the return on investing in governed context is compound—it simultaneously improves Agent accuracy, the auditability of data analysis, and cross-departmental collaboration efficiency. For Agent products, "precision" should be redefined as "reproducibly correct within governed context" rather than a high score on a single evaluation.
Finance and underwriting giving two answers isn't scary. What's scary is an Agent choosing one for you without telling you why.
One-line Summary
Genie One's accuracy problem is fundamentally a lack of context governance, not insufficient model capability.
Paragraph Summary
Using Genie One as a case study, the article explores why curated Agents' accuracy collapses in cross-departmental scenarios. Finance and underwriting give opposite answers to the same question—not because of data fabrication, but because each operates in a different data context. Governed context comprises three indispensable layers: definition consistency, permission visibility, and version timeliness. Curated Agents excel in single-department evaluations but default to familiar conventions across domains and don't proactively question their context, causing governance defects to be amplified as-is. The article proposes four engineering directions: unified semantic layer, context traveling with queries, output with lineage, and forced alignment for cross-domain questions. It also notes that Agent competition is shifting from model capability to context engineering quality, and that "precision" should be redefined as being reproducibly correct within governed context.
Related articles

Anthropic Founding Team in Their Own Words: Why We Bet on AI Safety
Anthropic's founding team explains their original motivation: from theoretical physics to AI safety, why they chose a public benefit corporation, how clumsy early models were, and the trade-offs behind launching Project Glasswing after the Mythos model revealed risks.

Black Holes Aren't Holes: Using the Trampoline Analogy to Explain Gravitational Spacetime Curvature
Black holes are not empty holes but mountains of matter. A video uses the trampoline and anvil analogy to explain how gravity bends spacetime, revealing the misleading origin of the name and helping readers build more accurate physical intuition.

Genie ZeroOps Public Beta: How Autonomous Agents Self-Heal Data Pipelines
Genie ZeroOps enters public beta—an agent that autonomously detects, investigates, and fixes data pipelines. We analyze its core capabilities, the 80,000 customer-issue experience behind it, and the real-world challenges of autonomous agents in data operations.