How CubiCube Uses Pegasus for Optimized Model Selection
CubiCube, the modular deployment platform, leverages the Pegasus Orchestrator and KnowYourModel registry to deliver intelligent, secure, and auditable AI operations to its users.
The Challenge
CubiCube enables creators to build and deploy web applications using modular "cubes"—pre-built components for databases, authentication, payments, and more. As the platform grew, the team faced a critical question:
"How do we give our AI assistant the power to help users across all these domains without creating a security nightmare?"
A monolithic AI assistant with access to everything—databases, payment keys, deployment credentials—would be a single point of failure. One compromised prompt could expose the entire system.
The Solution: Pegasus Orchestrator
Instead of one all-powerful agent, CubiCube adopted the Pegasus Orchestrator—a master coordinator that delegates tasks to 10 specialized sub-agents, each with its own scoped credentials.
Key Benefits
Least-Privilege Security
Each sub-agent only has access to the credentials it needs. The database agent can't touch payment keys. The email agent can't access storage buckets. A compromised sub-agent is contained to its domain.
Optimized Model Selection
Pegasus queries the KnowYourModel registry to select the best model for each task. Complex reasoning tasks get Claude Opus. Quick code generation uses Sonnet. The registry's usage-proof voting ensures recommendations are based on real-world performance.
Cryptographic Audit Trail
Every sub-agent invocation generates a usage receipt—cryptographically signed proof of what ran, when, and for how long. Agent definitions are Git-verified, linking each action back to its source code. CubiCube can answer "what did the AI do?" with verifiable evidence.
Standards-Based Interoperability
All agents follow the A2A Protocol—an open standard for agent-to-agent communication. This means CubiCube's orchestrator can discover and work with agents from other systems, and external tools can integrate with CubiCube's agents seamlessly.
Architecture Overview
Live Skill Registries
Each sub-agent capability is managed through its own registry. The orchestrator queries these registries at runtime to discover the top-ranked agent for each task. Explore the live registries:
The Master Orchestrator registry allows admins to swap orchestrators without code changes—the same pattern used for all sub-agents.
Deep Dive: Security White Paper
Coming SoonOur upcoming white paper, "Transparency vs. Opacity: A Comparative Security Analysis of Agentic AI Architectures", explores the security implications of orchestrator-based systems in depth—including comparisons with other approaches and OWASP Agentic Top 10 compliance.
Ready to explore the KnowYourModel Trust Registry?