From agent soup to proper software design: Mellea puts developers back in control
summary
Mellea is a Python generative‑AI library that claims to replace monolithic prompts with instructions not prompts, enforces task decomposition and information hiding, and adds compositional contracts between modules. The talk gives concrete patterns (not just hype): IVR (instruct validate repair) for automatic validate/repair loops, measurable effects like ~20‑point improvements on agent satisfaction, and cites industry failure estimates (MIT ~95% pilots fail, surveys ~80–90%). Works with existing agent frameworks, aims to lower inference costs and make fault models manageable.
AI-written summary · about
chapters (7)
- 0:00Agents disrupting software
- 1:30Why AI agent projects fail
- 3:59Prompt engineering as software
- 8:35Introducing Mallea
- 10:00Control flow not left to model
- 12:27Handling unreliable AI
- 14:11Closing with Mallea
related talks
01h 24mGenerative Programming with Mellea: from Agentic
50mKEYNOTE: The agentification of software has a UX problem
32mProgramming language evolution in the AI era
19mDon't Build Slop (4 Levels of AI Agent Maturity)
59mFrom Pair to Peer Programmer
21mBuilding Agents (the hard parts!)
27mAgentic Engineering: Working With AI, Not Just Using It
29mThe Continued Developer Evolution