MCP & agent infrastructure
In-depth technical writing on building production AI: how we design governed MCP servers, connect products to LLM clients, and orchestrate multi-agent systems that hold up under real load.
Where Your LLMs Run: Cloud APIs, Routers, and Self-Hosting
A practical guide to choosing where your models run: frontier cloud APIs, routers like OpenRouter, dedicated inference hosts like Fireworks, and self-hosting open models with Ollama, all behind one swappable interface.
Read paperOAuth 2.0 for Remote MCP Servers: Connecting Your SaaS to ChatGPT and Claude
What it actually takes to let ChatGPT or Claude connect to your product as a remote MCP server: dynamic client registration, PKCE, identity-provider-backed flows, discovery metadata, and mapping an outside user to the right tenant.
Read paperMulti-Agent Orchestration in Production: Routing, Guardrails, and Escalation
Lessons from running multi-agent AI systems in production: classifier-based routing with sticky sessions, bounded tool loops, per-agent and per-trust-level tool policy, confidence-driven human escalation, and budget enforcement.
Read paperGoverned MCP Servers: Safely Exposing Enterprise Data and Tools to LLMs
A practical architecture for exposing internal data and tools to LLM clients through the Model Context Protocol: per-agent grants, read/write tool scoping, SQL safety, approval workflows, and full audit trails.
Read paper