Composable Architecture for AI Systems: APIs, MCP, Skills, Sub-Agents, and Wor…
summary
Claim: hybrid systems should push repeatable LLM behavior into deterministic components and APIs so runtime AI stays minimal. Uses: MCP to let models invoke deterministic tools (99.9% of MCP tools run deterministically), skills and sub‑agents for modularized model work, and workflows to crystallize sequences (example: replace an LLM stitching 20 endpoints with a 20‑step deterministic workflow). Rationale: LLM calls cost orders of magnitude more (trillions× compute) and vary across model updates; moving repeatable steps to code reduces cost, unpredictability, and operational risk.
AI-written summary · about
chapters (9)
- 0:00Hybrid AI and code solutions
- 2:34AI not right for every problem
- 5:41Deterministic vs probabilistic
- 10:43Software architecture for hybrid AI
- 17:53Trade-offs: deterministic vs probabilistic
- 23:35Choosing AI integration mechanism
- 27:54AI changes API management
- 29:41Minimize AI for reliability
- 30:24Q&A
related talks
20mBuilding Deterministic Infrastructure for Non Deterministic AI Agents
30mComposable AI: Building Next-Gen AI Agents with MCP
27mThe Future is Non-Deterministic
32mFrom Monolith to AI Agent Modernizing Java Systems with MCP
50mFrom Prompt Hacking to Architectural Determinism: Engineering Reliable GenAI Systems
31mAgentic AI architecture patterns
34mAI Agents: Transforming Autonomous Decision-Making
29mArchitecting the AI Bridge: Integrating LLMs and DMN With Drools, Spring Boot, and the MCP Protocol