The Postmodern Data Stack: Where AI and Data Become One 

Theory Ventures VC Tomasz Tunguz recently argued that the postmodern data stack is AI. Tomasz is right; we’re in the midst of a transformation. At Monte Carlo, we’ve spoken to dozens of data and AI leaders across enterprise organizations and many of them agree: we’re expanding from a world where data systems primarily serve dashboards and analysts, toward one where they will also power intelligent, autonomous systems operating in real time.

These new players in the postmodern data stack, the AI agents, are changing the ways we build and observe our systems. But if AI truly defines the postmodern data stack, then it cannot be treated as a downstream consumer of data. 

In today’s production environments, AI systems don’t simply read data and return insights. They query operational layers, assemble context across structured and unstructured sources, reason probabilistically, invoke tools, and interact directly with humans. 

As agents work within the data layers, they generate their own telemetry as a byproduct of their analysis, decision-making, and interactions. This telemetry is a critical part of the input/output loop that we can use to think about the postmodern data stack in abstract. 

The more you think about it this way, the more you see how AI doesn’t sit on top of the data stack; it lives inside it. 

The input/output loop is the new system

Every AI agent operates through an input/output loop:

• Inputs: data, documents, embeddings, system state, user context
• Processing: context assembly, reasoning, tool execution
• Outputs: decisions, responses, and actions

The act of querying data, interpreting it, and generating out an output is in itself a production process. It creates its own telemetry: prompts, traces, tool calls, latency, evaluations. These are not exhaust artifacts; they are first-class signals that determine whether the system behaves correctly. And so, the process of generating an output generates even more data to be observed and analyzed… hence the loop!

Correct outputs are the result of both data quality and agent behavior. Incorrect outputs are rarely “just a data problem” or “just a model problem;” they are system failures across the loop.

This is why AI cannot be treated as a consumer of the data stack. It actively participates in, and reshapes, the data lifecycle. And so, like the data itself, AI must be observed. 

From trust contracts to system reality

In Redefining AI Agent Trust for Production, we framed trust as an input/output contract: reliable inputs, verifiable outputs, and accountability when expectations are violated.

But between inputs and outputs sits a probabilistic execution layer: the agent itself. Without visibility into how an agent assembled context, which data it relied on, how it reasoned, and what actions it took, trust remains theoretical. The newly-generated telemetry that feeds the loop is what actually gives us this visibility; we just have to now observe it. 

Observing data alone, like observing AI behavior alone, is insufficient. Trust only emerges when both are observed together.

Unified observability for the postmodern stack

If AI lives inside the data stack – querying it, interpreting it, and producing new data artifacts – then observability must live there too. 

But not as a bolt-on monitoring tool. Observability for the postmodern stack must act as a unifying control plane that connects:

• Data quality, freshness, and lineage
• Agent execution paths and telemetry
• Model behavior and output evaluation

Tunguz implies as much in his piece; a postmodern data stack demands unified observability that treats data and AI behavior as inseparable. 

Effective tools and workflows will recognize that data and agents are interconnected within this input/output loop that represents the new system, and will strive to offer full visibility into how they operate together. 

Envisioning the new architecture

Reliability in the postmodern data stack is no longer just about pipelines completing on time or dashboards refreshing correctly. It’s about whether autonomous systems make correct decisions under changing conditions.

This requires visibility into all aspects of the loop: the data itself, the agent’s interactions with that data, and the outputs those interactions produce. 

So what will the architecture look like in this new, postmodern data and AI stack? 

At Monte Carlo, we’ve identified an emerging four-layer architecture that encapsulates the following: a data build layer, a semantics and contextual layer, a multi-platform agent build layer, and an observability layer. 

We’ll soon be sharing a deeper exploration into this architectural framework and what it ultimately means for data and AI leaders building and implementing their organizational strategies.  

Ultimately, conversations about architectural design, governance, trust, and the many aspects of maintaining data excellence in today’s environment hinge on the truths we discussed in this post. AI is intrinsic to the postmodern data stack – much like an organ within the body. 

Observability that’s built for this new landscape can only succeed when it respects and works with this interconnectedness. That’s how we can nurture a strong input/output loop of trust and reliability at the scale we need to unlock the real value of this technology. 

The bottom line

If AI agents live inside the data stack, then trust depends on observing the entire input/output loop. This includes the data they rely on, the decisions they make, and the outputs they produce.

Monte Carlo helps teams do exactly this, empowering organizations to observe the entire Data + AI stack so autonomous systems can be deployed safely and at scale. 

To see the power of data + AI observability in action, schedule a demo with us. 

You can learn more about Agent Observability at Monte Carlo here.

Our promise: we will show you the product.