The Agentic Shift- Are We Reasoning, Or Just Retrieving Smarter?
summary
Argues that 2025 is the era of agents not magic: RAG systems give context while agentic frameworks (using tool APIs and ReAct patterns) generate their own multi-step plans, spawn sub-agents and iterate with default max_iter ~3–5. Claims Piramal sees ~92% of new code authored by AI, notes Claude 4’s 1M-token memory but warns memory ≠ mind, shows reasoning models collapse on large Tower-of-Hanoi puzzles and are vulnerable to token-bias/prompt-injection. Recommends micro-pilots, guardrail-as-a-service and always verify outputs; frames tradeoff as Fast Thinker (RAG) vs Slow Thinker (agentic).
AI-written summary · about
chapters (22)
- 0:00Agentic Shift introduction
- 2:32AI agents and reasoning models
- 6:23Three AI evolution shifts
- 9:13Context in AI responses
- 13:09How agentic systems work
- 16:17Illusion of AI thinking
- 18:22Human vs AI reasoning
- 21:15Testing AI with Linda Problem
- 24:43AI decision-making biases
- 29:00AI reasoning limitations
- 31:50Overthinking and token bias
- 33:13Agentic engineering vs AI
- 35:41Guardrails and data security
- 38:55Q&A
- 40:00Data security in banking AI
- 42:11Bias in AI-driven QA
- 43:57Can AI reason like humans?
- 44:43Ethical accountability in AI
- 45:25AI creativity and hallucinations
- 46:52AI in non-software roles
- 49:02Skills for testers
- 50:06Closing Remarks and Key Takeaways
related talks
28mAgentic AI
27mEmerging Patterns Agentic AI
19mRedefining Intelligence: How Reasoning Is Re-Shaping AI in 2025
20mReimagining InnerSource for the Agentic AI Era
33mAsk Me Anything: AI Agents, Skills, Career & Growth
17mRethinking AI Safety in the Age of Agentic Systems
28mAgentic Enterprise - What your CEO must know about AI
46mIntelligence as Infrastructure: What Developers Must Learn to Build with AI