• The MOB
  • Posts
  • Everything You Need to Know About AI Agents

Everything You Need to Know About AI Agents

and how capitalize on it right away

Author

Ethan Sperry
June 15, 2025

Agentic AI = Language models that can reason, plan, and interact with external environments through iterative processes, moving beyond simple text-in/text-out interactions.

FOUNDATIONAL KNOWLEDGE

Language Model Training Pipeline

  1. Pre-training: Models learn from massive internet datasets to predict next words

  2. Post-training: Two critical steps make models usable:

    • Instruction following: Learning to respond to specific prompts properly

    • RLHF (Reinforcement Learning with Human Feedback): Aligning outputs with human preferences

Current LLM Limitations

  • Hallucination: Generating incorrect information confidently

  • Knowledge cutoffs: Can't access recent information

  • No attribution: Can't cite sources for claims

  • Data privacy: Can't access proprietary/private information

  • Context limits: Finite window for processing information

BRIDGING SOLUTIONS

RAG (Retrieval Augmented Generation)

  • Process: Pre-index documents → Convert to embeddings → Store in vector database → Retrieve relevant chunks for queries

  • Benefits: Accurate, attributable responses using up-to-date information

  • Use case: When you need models to work with proprietary or recent data

Tool Usage (Function Calling)

  • Capability: Models generate properly formatted API calls that external software executes

  • Evolution: From describing actions to actually performing them

  • Example: Instead of explaining how to check weather, actually calling weather API and returning real data

THE FOUR AGENTIC DESIGN PATTERNS

1. Planning

  • Core idea: Break complex tasks into manageable subtasks

  • Process: Ask LLM to create execution plan → Work through each component systematically

  • Why it works: Decomposition consistently outperforms monolithic approaches

2. Reflection

  • Core idea: Self-improvement through iterative critique

  • Process: Generate initial response → Analyze own output → Identify improvements → Produce refined result

  • Best for: Code refactoring, writing improvement, quality enhancement

3. Tool Usage

  • Evolution: From simple function calling → sophisticated environmental interaction

  • Capabilities: Code execution in sandboxed environments, complex API integration, multi-system coordination

  • Impact: Transforms models from information providers to action-oriented agents

4. Multi-Agent Collaboration

  • Structure: Specialized agents for specific domains coordinated by central orchestrator

  • Benefits: Improved performance through specialization, better maintainability

  • Example: Smart home with separate climate, lighting, and security agents

KEY INSIGHTS

The ReAct Paradigm

  • Reasoning (chain of thought) + Action (tool usage)

  • Combines thinking processes with real-world interaction capabilities

Capability Ceiling Breakthrough

  • Same language models can tackle problems beyond their apparent limits

  • Agentic patterns unlock hidden capabilities in existing technology

  • Success comes from better usage patterns, not just better models

Customer Support Example Workflow

  1. Receive refund request

  2. Check company refund policy

  3. Retrieve customer information

  4. Verify product details

  5. Make refund decision

  6. Execute appropriate action

A FEW EXAMPLES

Software Development

  • Analyze codebases and identify bugs

  • Propose and test fixes in isolated environments

  • Submit pull requests autonomously

  • Function as AI developers

Research & Analysis

  • Systematically gather information from multiple sources

  • Synthesize findings into comprehensive reports

  • Generate insights that would take humans hours to compile

Task Automation

  • Handle complex multi-step workflows

  • Coordinate between different systems

  • Adapt to changing requirements dynamically

IMPLEMENTATION STRATEGY

  1. Start simple: Master basic LLM usage first

  2. Experiment safely: Use playground environments before production

  3. Build incrementally: Begin with reflection, add tool usage, advance to planning

  4. Implement logging: Systematic evaluation and monitoring from day one

Success Factors

  • Proper prompting strategies remain foundational

  • Systematic approach to evaluation and testing

  • Understanding when to apply which patterns

  • Building robust error handling and fallback mechanisms