Foundational research on agentic AI, edge computing, and privacy-first systems. Our work advances the field through rigorous analysis, performance benchmarking, and security innovation. All research is open to the community and drives industry standards.
Our work advances the field of agentic AI, edge computing, and privacy-first systems. All research is open to the community and drives industry standards.
Our flagship technical whitepaper detailing the VSSP protocol, Adversarial Governance, and the transition from stateless bots to high-context digital societies. Essential reading for understanding VoidCat RDC architecture.
Our foundational research posits that agent performance stems from intelligent context and tool governance, not raw model size. This comprehensive whitepaper covers tool-RAG hybrids, context offloading, definition compression, and real-world benchmarks showing 40% accuracy improvement and 60% token reduction.
Analysis of tool routing strategies for complex queries. Compares RAG-first vs Tool-first approaches across 500+ evaluation cases with performance/accuracy trade-offs. Establishes decision framework for choosing optimal strategy per query type.
Analysis of ~2,000 vulnerable MCP servers identified in June 2025 security research. Documents authentication gaps, over-permissioning issues, and prompt injection vectors. Details our OAuth 2.1 and RFC 8707 implementation as industry solution.
Deep technical analysis of deploying reasoning engines on edge networks. Covers optimization techniques (model quantization, context pruning), cost models, and deployment patterns across 330+ Cloudflare cities achieving P99 <300ms.
Framework for loading tools only when needed, reducing context pollution and enabling 100+ tool definitions without performance degradation. Includes evaluation harness, benchmarks, and production deployment patterns.
Introduction to our proprietary multi-agent orchestration platform for coordinating reasoning, execution, and verification agents. Covers state management, failure recovery, inter-agent communication, and real-world applications.
Most AI platforms assume bigger models = better results. We posit the opposite: intelligent context management is the differentiator.
Route user intent through two paths: (1) RAG for retrieval-first tasks, (2) Tool-calling for external integrations. Choose optimal path per query.
Load tools only when needed, not all at once. Reduces context pollution. Allows 100+ tool definitions without degrading performance.
For long-running tasks, offload context to storage. Summarize intelligently. Enables multi-hour agents without token explosion.
Compress verbose OpenAPI specs into minimal descriptions. Preserve semantics. Reduce tokens by 70% while maintaining accuracy.
Every technical decision at VoidCat is evaluated across eight quality dimensions. FLAGS ensures we balance innovation speed with production reliability:
Grading: Each dimension scored 0-10. Target across FLAGS: ≥8/10 average. Code failing <6/10 on any dimension blocks release. PR reviews explicitly evaluate code against FLAGS dimensions.
Features must pass five quality gates before reaching production. Each gate is a checkpoint where work can be approved, rejected, or sent back for revision.
Question: Is this worth building?
Approval: Founder signs off. No development begins without Gate 0 approval.
Question: Is the code production-ready?
Approval: Code review checklist passed. CI/CD gates green.
Question: Does it work with dependent systems?
Approval: Integration test report signed by QA lead.
Question: Is it ready for real users?
Approval: Product lead certifies readiness. On-call acknowledges runbook.
Question: Is it stable in production?
Graduation: Feature marked stable. On-call handoff complete. Metrics baseline for future regression detection.
Enforcement: Each gate is a hard stop. Work cannot proceed to next gate without explicit approval. Exceptions require founder waiver (documented).
Security is built into architecture, not bolted on. These baselines apply to all systems:
VoidCat's research foundations are shared with the community:
