Last week in New Orleans (aka The Big Easy), Current 2025 delivered on expectations. Confluent brought ML functions and streaming agents to Flink; incremental materialized views that keep context up-to-date for AI agents; Confluent Private Cloud to meet on-premises requirements for their SaaS offering; and ... a Kafka proxy.

The Conduktor team at our Current booth.

There is a quiet democratization of the proxy pattern across the streaming ecosystem. For years, streaming vendors competed on performance. But now, many are building the same thing: a control plane on top of streaming. Confluent announced its gateway: a transparent proxy layer for cross-cluster routing, based on Kroxylicious, the CNCF Kafka proxy from Redhat. Kong and Gravitee entered the space recently. 

Today, everyone is shipping a proxy.

When we launched the first enterprise-grade proxy for Kafka back in 2022, the idea was radical. It was built to protect Kafka from bad practices and govern heterogeneous ecosystems at scale, providing the first dedicated control plane for streaming.

2022: "Single point of failure!" "Latency!" Now, that same idea has become the new architectural layer for data streaming, not a nice-to-have but an integration layer on top of the streaming infrastructure.

This is the great confluence: not just of tools or vendors, but of philosophy. The industry has realized that governance, not infrastructure, is a bottleneck (especially with AI models that need ALL the data!). The proxy is an architectural control point that removes this bottleneck.

The Industry Shift

These are the challenges everyone is focusing on:

1. Governance versus velocity: How do you maintain control without introducing a bottleneck?

2. Compliance versus delivery speed: Why does security still mean slowing everything down?

3. Visibility versus cost: Who’s using what, and what is it actually worth?

These aren’t new problems. What’s new is that the needs have multiplied: not just in data volume, but in pattern complexity, heterogeneity of infrastructure, and scale (cross-regions etc.).

Streaming is mainstream, well understood, and easy to get quick value from (but 'quick' does not mean 'good'). There is a growing need for automation: pattern detection, policy enforcement and adaptive governance. Less humans in the loop.

We’ve escaped the analytics silos ... only to stumble into streaming silos: multiple clusters, cloud boundaries, lack of infrastructure flexibility.

That's the confluence: Streaming at scale demands a unified control plane. Enterprises are discovering they need it.

Governance at Speed

Governance was designed to protect us and our data. Somewhere along the way, it became an anchor holding your team back, rather than an enabler helping you get things done. The results: unused tooling piling up, a lack of coordination, and a lack of 'governance culture.' 

Just ask yourself: do your application teams actually use Collibra, or is it just your isolated governance team?

The proxy introduces a connective layer between heterogeneous systems, decoupling control from infrastructure. Lack of governance is detected and can become code.

When we built Conduktor's proxy in 2022, it wasn't born from theory. It was to cover a large gap in the market. Watching everyone struggle to make Kafka usable and safe at scale, without an army of professional services to fine-tune it and educate your teams. It was acceptable to move fast, break things quickly, and keep applying patch after patch while everyone kept attacking Kafka.

The proxy changed that. Sitting between apps and clusters, it became the single enforcement layer: managing access, encryption, auditing, and masking without touching producers or consumers. The goal was simple: keep Kafka open, but make it governable.

What started as a technical conversation ("single point of failure", "latency") evolved into business conversations: a way to enable and spread streaming faster.

The Maturity Correction

Most organizations aren’t failing because their streaming tools are bad: they’re failing because their operating models are.

Kafka promised freedom — the freedom to move data. But learning how to use it well: understanding how it works, managing access, identifying the data flowing inside, all this requires organizational maturity. It’s not a tech problem.

As Stanislav Kozlovski's recent piece on streaming's plateau points out, we're seeing growth deceleration across the board. Confluent's cloud growth is slowing. Stream processing adoption remains niche (Flink generated just ~$14M ARR for Confluent in Q3, barely 1.25% of their business). The industry's not shrinking, but it's definitely not in hypergrowth mode anymore.

This is a symptom of a maturity correction. The infrastructure is fine; the governance isn’t. Enterprises (or their providers) have mastered running Kafka, but are not using it well.

The proxy pattern (combined with true self-service capabilities) is how you move past this bottleneck. It's the natural correction to years of over-centralized complexity. It’s an evolution from “infrastructure-first” to “control-first”. Governance becomes dynamic, enforced at the edge, and has "shifted left". Streaming architecture can now balance innovation with discipline.

From Data in Motion to Decisions in Motion

If 2020 was the decade of “data in motion”, 2025 is the dawn of decisions in motion.

The Current keynote hit hard on agent-driven workflows and AI-native architectures. Confluent Intelligence, Streaming Agents, and Real-Time Context Engine all position data streaming as the foundation for the next wave of AI applications. 

Is this going to happen, or not? Is streaming the backbone for agent-driven workflows?

As enterprises adopt agentic systems, the need for real-time, secure, on-demand access to data will only grow. Agents can't wait for batch jobs, they can't wait for slow access approval. They need data access and context now: streamed, filtered, and audited.

I'm personally a big fan of Kafka's new queue semantics for this. Queues provide concurrency and reliability; proxies provide access, trust and governance. Together, they form the nervous system of AI: a live data fabric where every decision is observable, auditable, and compliant by design. I spoke about this during my talk at Current, using a chess game as a metaphor for distributed decision-making with LLMs. Same thing.

Most enterprises aren’t there yet but AI is forcing data governance to grow up.

Incoming: MCP over Data?

Imagine a control plane that doesn’t just enforce policy but understands intent. An agent or application doesn’t ask where the data lives — it asks what it needs: “customer transactions,” “fraud alerts,” “active sessions.” The MCP layer, built on top of the proxy, resolves the source dynamically, applies transformations, masks sensitive fields, and enforces access — without the client knowing where the data resides.

This is the natural extension of what proxies already do. They’re the connective tissue between applications and streams, with full visibility into requests, flows, and access patterns. They already see everything. The next step is teaching them to reason.

When that happens, data stops being a location problem and becomes a context problem. The proxy becomes a gatekeeper, translator, and policy engine in one — the foundation for MCP: dynamic, least-privilege, on-demand access for humans, services, and AI agents alike.

As data architectures fragment across clouds and AI agents multiply, static pipelines will collapse. The next generation of systems will demand adaptive control, policy that travels with the request, not the infrastructure.

Where We Go From Here

As Yaroslav Tkachenko observed in his Current recap, there's skepticism about whether we're in an AI bubble and that streaming is trying to escape there. I share that skepticism: working with tons of customers, this is not really where the problem is today. There is still a huge need to simply stream data reliably, govern it precisely, enable self-service safely, and innovate faster without losing control. We talk about AI on one side, but the reality is struggling with cost, partitions, and understanding how data flows (or should flow).

Since 2022, our mission has been to help customers get more out of what they already have. By making existing systems easier to use, safer to share, and simpler to govern, we enable teams to move data freely and securely to maximize infrastructure usage and turn streaming into an engine of innovation.

Because in streaming’s great convergence, one truth stands: Control is the new compute. And the proxy is where it begins.