Build Reasonable Software
summary
Concrete claims: FastMCP v3 redesign compresses a 2,500‑line PR to 400 lines by reusing abstractions and replaces brittle feature bloat so it actually works, adding code mode yields a 90% reduction in tokens and latency for complex MCP servers, Prefab composes front ends in Python and streaming Python gives about a 70% token saving, and a proposed directed agentic graph makes agent outcomes observable, serializable, and composable for safer orchestration.
AI-written summary · about
chapters (13)
- 0:00Building reasonable software
- 2:53Pythonic software meaning
- 6:24Reasonable software in AI age
- 8:28FastMCP 2 to 3 redesign
- 11:58Code mode and token efficiency
- 13:24Introducing MCP apps and Prefab
- 17:15Prefab’s Pythonic UI
- 20:01Reasoning software
- 22:54Control vs autonomy in AI
- 24:37Agentic graphs for governance
- 27:47Agent workflows as DAGs
- 31:58Control vs agent harnesses
- 32:38Q&A
related talks
55mYour MCP Server is Bad (and you should feel bad)
19mStop babysitting your agents...
12mMCP and the Dawn of Agentic Intelligence
22mHow Block Operationalized MCP At Scale
53mIntentional Code - Minimalism in a World of Dogmatic Design
27mHow We Built Our First MCP Server: Lessons From the Trenches at Hud
15mFull Spec MCP: Hidden Capabilities of the MCP spec
12mDesign-to-code and beyond: inside Coinbase’s MCP ecosystem