Hands-on with CAIPE: Building an open source, multi-agent system for platform engineering

Platform engineering is the pillar of modern software development, giving developers the tools, environment, and automation they need to ship quick and reliable code. 

In the past decade, the role of platform engineering has grown increasingly multifaceted. Watershed moments like the rise of Kubernetes and containerization and the explosion of cloud native architecture have made engineers responsible for a vast ecosystem of intricate tools and technologies.

For most teams, the work required is incredibly demanding. It means long support queues, constant problem solving, and context-switching between disparate systems.

Like many teams, the Outshift by Cisco platform engineering team experienced these hurdles. It left little time for creative, high-value work. The team looked at their operations and identified how agentic AI could help them work smarter.

CAIPE: An open source, multi-agent system for platform engineers

Community AI Platform Engineering (CAIPE) is Cisco’s open source, multi-agent system, built to automate platform operations and give engineers relief from repetitive, manual tasks. 

Supported by the Cloud Native Open Source Ecosystem (CNOE) Agentic AI Community, CAIPE provides a secure, scalable, persona-driven reference implementation with built-in knowledge base retrieval.  This streamlines platform operations, accelerates workflows, and fosters innovation. It integrates seamlessly with Internal Developer Portals like Backstage, VS Code, and CLI tools, enabling frictionless adoption and extensibility.

CAIPE originated as JARVIS, an internal Cisco project exploring how to offload operational work using AI. It grew into a framework where multiple specialized agents work together under a supervising agent, each focused on a distinct domain. CAIPE delegates work to ensure a cleaner architecture and better scalability. 

It doesn’t force every tool and capability into a single, overburdened model. Instead, each agent works on specialized tasks, such as deploying to Kubernetes, updating Jira tickets, retrieving knowledge from a documentation database, or committing code to GitHub.

“We realized that cramming all tools into a single agent doesn’t work,” Sri Aradhyula, a Platform Engineering Architect at Outshift, says. “The supervisor-and-sub-agent pattern lets us keep large language model (LLM) context windows small and helps the supervisor orchestrate more complex workflows.”

As an open source project, CAIPE embodies Outshift’s commitment to transparency, collaboration, and community-driven innovation. By contributing to the CNOE Agentic AI Community, we’re fostering an ecosystem where engineers, developers, and organizations can come together to shape the future of platform engineering.

CAIPE allows engineers to customize, extend, and interconnect their own AI agents using shared standards like A2A and MCP, integrating components from the AGNTCY project, such as Agent Identity, Agent Directory, and SLIM (Secure low-latency interactive messaging) communication protocols.

By using CAIPE, Outshift’s team has automated around a third of its internal platform engineering tasks. They’ve reduced the average incident response time from hours to seconds and reclaimed time for strategic work. 

CAIPE hands-on lab: Creating and deploying secure, enterprise-ready multi-agent systems

The Outshift Platform Engineering team hosted a workshop to help the wider Cisco community run, modify, and deploy their open source, secure, enterprise-ready, distributed multi-agent system.

Each attendee stepped into the role of an engineer of a fictional Mars colony. Their mission was to establish Mission Control using CAIPE.

During the lab, participants connected to the workshop’s pre-built environment and followed a set of guided missions. Each mission took a single agent to a production-ready, multi-agent deployment. They learned the fundamentals of CAIPE’s multi-agent system architecture, best practices for agent orchestration, and how to monitor and trace agent behavior.

The hands-on lab drew strong engagement with 168 active attendees, and over 100 attendees completing the lab mission. Every participant who finished missions was awarded with digital badges to celebrate their success.

From ReAct agents to production-ready multi-agent systems

The workshop started with a simple ReAct-pattern agent, showing how an AI model can reason and act through tool calls, process results, and provide a final answer. Attendees built and ran standalone agents in both HTTP and SDIO MCP modes before moving on to one of CAIPE’s defining patterns: a supervisor agent managing domain-specific sub-agents. 

Another mission introduced a retrieval-augmented generation (RAG) agent that ingested Mars research from Wikipedia and a GitHub agent that formatted the research into a Markdown mission report that committed it directly to a repository. CAIPE’s supervisor orchestrated these cross-agent interactions over AGNTCY SLIM.

The last module brought everything to production scale. Using IDP Builder, participants deployed CAIPE with Backstage and ArgoCD integration and then applied distributed tracing using OpenTelemetry, and AGNTCY protocols to monitor and measure agent decision-making.

Join the open source CAIPE community 

From an internal project to an open source platform, CAIPE shows how platform engineers can bring AI agents together to increase the scale and pace of their work. 

You can start using CAIPE now. Visit the public GitHub repository and documentation to try it for yourself.

We look forward to seeing your contributions to the community building the future of AI platform engineering and plan to have an external lab in the coming months!