Our Principles
How we engage.
Foundation first
We build the clean, governed data foundation before the dashboards. It's the unglamorous step most projects skip — and the reason ours hold up.
Fast on lightweight cloud
Efficient cloud architecture gets you to value in weeks, not the multi-quarter slogs the big firms quote.
Written by humans
We use the best tools available — including AI — but the judgment and accountability are human. Real analysts sit alongside the technology.
Measured by outcomes
We judge the work by what it changes — acquisition, retention, decisions made — not the features shipped.
The Delivery Path
From scattered sources to dashboards people trust.
Every engagement follows the same path. We unify what’s scattered, govern it, then build the models and dashboards on a foundation that lasts.
Sources
Apps · APIs · files
Building Production AI
A six-layer framework for agentic systems.
When we design AI systems for clients, we map every component to six layers before writing code. The implementation stack changes from engagement to engagement; the layers don’t. It’s how an agent stays accountable, observable, and under human control in production.
Guardrails, escalation, human-in-the-loop, kill switches
Every prompt, retrieval, decision and outcome logged and measured
Executes against systems of record — observable, reversible, gated
Multi-step planning, tool selection, uncertainty handling
What the agent knows at the moment it acts, surfaced at runtime
Who the agent is, what it can see and do — scoped and audited
Read top to bottom: Governance keeps humans in control, Identity is the foundation every action is attributed to.
In practice, Context means more than “the agent can see your docs.” Every knowledge source it works from carries a trust signal — approved, in progress, or noise to ignore — so the system builds from what’s true, not whatever’s newest in the folder. And Governance isn’t a settings toggle: any specification an agent will build from gets a named human’s sign-off before work starts, with a defined process for what happens when that spec changes mid-build.
We’ve also learned the less glamorous half of Observability: a technically accurate log of what an agent did isn’t the same as an answer a non-technical stakeholder can act on. Every AI-driven status update we produce gets translated into what changed, who owns the next step, and what decision it’s waiting on.
Agentic Workflow FAQ
What is an agentic workflow?
An agentic workflow is a process in which an AI agent plans multi-step work, selects tools and acts on systems of record, while humans keep control at defined points. Cruz Street designs every agentic workflow against its Six-Layer Agentic Workflow Framework: Identity, Context, Reasoning, Action, Observability and Governance. Each layer has a defined job, and each has a characteristic failure when it is missing.
What is Cruz Street's Six-Layer Agentic Workflow Framework?+
It is the framework Cruz Street uses to design agentic systems: Identity, Context, Reasoning, Action, Observability and Governance. Every component of a system is mapped to one or more of the six layers before any code is written. The layers stay the same on every engagement; only the implementation stack changes.
What does the Identity layer do in an agentic workflow?+
Identity defines who the agent is, what it can see and what it can do. Every action is attributable, scoped and audited. When this layer is missing, the failure is an agent operating on client data with no audit trail or access boundary.
What does the Context layer do in an agentic workflow?+
Context is what the agent knows at the moment it acts: system data, workflow knowledge and operational context surfaced at runtime, not baked into prompts. When this layer is missing, the failure is a generic model with no awareness of client operations.
What does the Reasoning layer do in an agentic workflow?+
Reasoning is how the agent decides: multi-step planning, tool selection, action sequencing and uncertainty handling. When this layer is missing, the failure is unstructured model calls with no guardrails or plan.
What does the Action layer do in an agentic workflow?+
Action is how the agent executes: it integrates with systems of record, produces artifacts and completes transactions. Actions are observable, reversible where possible, and gated by identity. When this layer is missing, the failure is side effects with no traceability or rollback path.
What does the Observability layer do in an agentic workflow?+
Observability means every prompt, retrieval, decision and outcome is logged, and performance, cost and business impact are measured continuously. When this layer is missing, the failure is a black-box agent that cannot be audited or operated in production.
What does the Governance layer do in an agentic workflow?+
Governance defines what the agent cannot do, what it must escalate, and how humans stay in control, through guardrails, schema enforcement, human-in-the-loop gates and kill switches. Failures fail loud, fail to a human and fail recoverably. When this layer is missing, the failure is a runaway agent with no escalation path and silent failures.
Does the framework depend on a specific technology stack?+
No. The six layers apply regardless of tooling choices. Cruz Street's implementation choices change from engagement to engagement; the layers do not.
