Atlas · GenAI 2026
AI Agent Design
Agent architectures (ReAct, Plan-and-Solve, self-correction)
conceptPeak: 2025Agent ArchitectureAI consensus: 3/3
Prerequisites
Agents work by calling tools iteratively — function calling is the atomic operation that agent architectures compose
Agent loops (ReAct, Plan-and-Solve) rely on carefully engineered system prompts, reasoning prompts, and reflection prompts
- mediumLong-Context Modeling
Multi-step agent workflows accumulate context (observations, tool results, thoughts) — context management becomes critical
Recommended reference
Yao et al. (2023) 'ReAct: Synergizing Reasoning and Acting in Language Models' — ICLR; the foundational agent paper. Pair with Anthropic (2024) 'Building Effective Agents' for production patterns
Notes from AI deep research
Anthropic Opus
Yao (2023) ReAct = fundament. Anthropic 'Building Effective Agents' = best practice. Rok agentow
OpenAI Deep Research
Wzorce: dekompozycja, delegacja [OA#27]
Google Deep Think
ReAct, Plan-and-Solve, samorefleksja [G#51]
Related skills
- ← is part of: Agent Memory Systems(3/3)
- ← is part of: Agent State Management(3/3)
- → is subcategory of: GenAI(3/3)
- ← is subcategory of: Code Execution Agents(3/3)
- ← is subcategory of: Computer Use AI(3/3)
- ← is part of: Human-in-the-Loop AI(3/3)
- ← is part of: LLM Function Calling(3/3)
- ← is an instance of: LangChain(3/3)
- ← is an instance of: Model Context Protocol(3/3)
- ← is subcategory of: Multi-Agent Orchestration(3/3)
- ← is part of: Prompt Engineering(3/3)
- ← is part of: Structured LLM Outputs(3/3)
- ← is part of: System Prompt Design(3/3)
- ← is subcategory of: Text-to-SQL(3/3)
- ← is part of: Agentic RAG(2/3)
- ← is an instance of: LangGraph(2/3)
- ← is an instance of: Pydantic AI(2/3)
- ← is subcategory of: Conversational AI(0/3)
- ← is subcategory of: Multi-Agent Systems(0/3)
- ← is subcategory of: Dialogue Systems(0/3)