Convert your guidelines into deterministic, enforceable rules wherever possible - then pair that with a solid AGENTS.md. That’s the recipe for high-ROI workflows. As @FactoryAI says: agents write the code, linters write the law.
https://t.co/0V4lTi4DkJ
Sir Demis: All biological systems are specialized to their respective niches but also generalized in the sense that they can evolve and learn as the environment changes. I see 2 primary differences in biological vs. man-made systems (intelligence seems to be an emergent property):
1. In biological systems, information processing and storage happen at the same place, unlike human-designed ones.
2. In biological systems the underlying hardware has a unique ability to evolve to optimize performance on two key goals: 1. Survival, 2. Reproduction.
You see countless examples in nature: locomotion patterns change due to changing environmental needs (from monkeys to apes) and the tail ceased to exist (through a mutation in the TBXT gene). There is long enough exposure to antibiotics, and the bacteria evolves resistance, and so on. On the other hand, our most intelligent systems (LLM-based apps) learn only through external memory implementations, which is so fragile and nothing like "learning".
This is what we must decode and try to replicate.
@ylecun
and then (dynamically) spins-off sub-agents to work on those specific tasks : how do you accomplish this in LangGraph? Does Agent SDK makes it easier to achieve this?
@langchain@OpenAI@OpenAIDevs
I love LangGraph and we haven't hit any limitations or found performance/scaling issues as application evolves. But if you have to handle a situtation wherein a domain specialized master agent breaks tasks down into sub-tasks...
Here's a great quote about 'Training' from a great leader, who has taken Indian cricket to new heights. As he turns 30 today, he keeps pushing the bench-mark even higher.
What's your thought on it? What are you keen on improving?
#training#leadership#development#success