Building safe, multi-agent AI systems in Redpanda Agentic Data Plane

At Redpanda, scaling agentic customer experience began in mid-2025 with Redleader, a single, prompt-driven agent operating across boundless Slack channels and customer interactions. If you attended Streamfest, you might remember our demo on designing Redpanda AI agents for customer replies. In short, Redleader proved that:

• AI could draft meaningful responses given the right data
• Triage could be automated to improve customer experience
• The scale program could leverage AI to generate measurable ROI
• Humans and AI needed to collaborate safely and with trust

It served as a foundation that delivered real value. However, as adoption increased we recognized it needed more structure and controls. So, we developed a revamped Redleader—a bounded, multi-agent system running on Redpanda’s Agentic Data Plane (ADP). 

In this post, we loop you in on what we changed in the product architecture and how it lets us operate Redleader at scale across our enterprise ecosystem.

Creating a multi-agent habitat

Redleader now consists of multiple bounded agents that collaborate, rather than a single agent that does everything. Responsibilities are granted to specialist agents:

• Concierge (ingestion)
• Records Custodian (durable state)
• Orchestrator (reasoning coordination)
• Sentiment Discriminator (risk classification)
• Activity Monitor (observability and learning)

To produce agent “collaboration,” we needed a multi-agent AI habitat (often called a “Substrate”) in which agents can exchange structured events, enforce governance, and learn continuously. In this case, the Redpanda Agentic Data Plane is our substrate.

Redpanda ADP brings together:

• Redpanda Streaming engine
• Redpanda Connect (integration and transformation layer)
• Agent orchestration components
• Observability streams
• Human-in-the-Loop control signals

Agents that “think” before they act

We needed agents not just to act but to coordinate, specialize, and improve the experience together. This system now produces continuous event streams:

• Inbound Slack messages
• Normalized records
• Routing decisions
• Escalation signals
• Human override reactions
• Audit metadata
• Learning exhaust

These are ordered, durable streams of intent, action, and reaction. Streaming serves as the agent-coordination mechanism and shared language of the ecosystem. Our business can track and trust the agents' operations, and the agents also communicate with each other through these events. We transformed from a capable agent into a coordinated system within a governed, autonomous ecosystem.

Ready to build?

Sign up for Redpanda Cloud today and give Redpanda Streaming and Redpanda Connect a go. If you want early access to Redpanda ADP, get in touch. 

In the meantime, you can dig into our Redpanda Connect workshop to learn the ropes and start building your own streaming pipelines.