Emerging Patterns Agentic AI
summary
Review of 120 projects across a 4,000-person org shows agentic systems cluster around three concrete patterns: code as action space where models generate runtime code to call deterministic APIs (use secured sandboxes like Firecracker microVMs to run it), Model Context Protocol (MCP) as a JSON spec not a transport protocol, and planners + memory (8K context limits, “lost in the middle,” Titan-style test-time memory and KBLM-style knowledge-in-model). Games (chess, AlphaGo, Diplomacy) validate search+learning; Anthropic suggests ~90% of cases only need workflows, not full agents.
AI-written summary · about
chapters (15)
- 0:00Introduction and context
- 2:08Games as agent testing environments
- 6:25The bitter lesson in AI
- 8:10What is an agent?
- 10:18Emerging agentic AI patterns
- 12:04Code as reasoning pattern
- 13:28Secured sandboxes for safety
- 14:12MCP as a standard
- 15:37Planning in agentic systems
- 17:43External planners vs LLMs
- 19:51Memory management in agents
- 20:55Context window limitations
- 22:57Runtime memory patterns
- 24:21Titan and KBLAN models
- 26:06When to use agents
related talks
14mThe 4 Patterns of AI Native Development
01h 59mEmerging Patterns for Coding with Generative AI
31mAgentic AI architecture patterns
49mAgentic AI Patterns
28mAgentic AI
47mAgentic AI Patterns
52mThe Agentic Shift- Are We Reasoning, Or Just Retrieving Smarter?
26mBuilding Agentic Workflows: Patterns for Orchestrating Intelligent Systems.