How we built an MCP integration, and what we learned along the way.
Why static system prompts aren't enough. How measurable behavioral dimensions create consistent agent identity across sessions and providers.
What happens when you give an AI agent real memory — with confidence scoring, time decay, contradiction detection, and a knowledge graph.
One user, multiple agent identities. How the sleeve abstraction gives AI agents context-appropriate personality, memory, and skills.
Not every AI call deserves a frontier model. How deliberate model-tier strategy keeps per-conversation costs under a penny without sacrificing quality.
We built MCP tools and agents refused to use them. The fix had nothing to do with the protocol.